research article dynamic self-occlusion avoidance approach
TRANSCRIPT
Research ArticleDynamic Self-Occlusion Avoidance Approach Based onthe Depth Image Sequence of Moving Visual Object
Shihui Zhang12 Huan He1 Yucheng Zhang1 Xin Li1 and Yu Sang1
1School of Information Science and Engineering Yanshan University Qinhuangdao 066004 China2The Key Laboratory for Computer Virtual Technology and System Integration of Hebei Province Qinhuangdao 066004 China
Correspondence should be addressed to Shihui Zhang sshhzzysueducn
Received 16 May 2016 Accepted 30 August 2016
Academic Editor Alessandro Gasparetto
Copyright copy 2016 Shihui Zhang et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
How to avoid the self-occlusion of a moving object is a challenging problem An approach for dynamically avoiding self-occlusionis proposed based on the depth image sequence of moving visual object Firstly two adjacent depth images of a moving object areacquired and each pixelrsquos 3D coordinates in two adjacent depth images are calculated by utilizing antiprojection transformation Onthis basis the best view model is constructed according to the self-occlusion information in the second depth image Secondly theGaussian curvature feature matrix corresponding to each depth image is calculated by using the pixelsrsquo 3D coordinates Thirdlybased on the characteristic that the Gaussian curvature is the intrinsic invariant of a surface the object motion estimation isimplemented by matching two Gaussian curvature feature matrices and using the coordinatesrsquo changes of the matched 3D pointsFinally combining the best viewmodel and themotion estimation result the optimization theory is adopted for planning the camerabehavior to accomplish dynamic self-occlusion avoidance process Experimental results demonstrate the proposed approach isfeasible and effective
1 Introduction
Self-occlusion avoidance refers to the fact that when the self-occlusion of visual object occurs the visual system changesthe camera observation direction and position by usingcurrent detected self-occlusion cue to observe self-occlusionregion further and obtains more information of the visualobject so as to accomplish the related visual task betterTherefore the main problem to be solved on self-occlusionavoidance is how to determine the best observation directionand position of the camera by utilizing the occlusion cue ofcurrent visual object
According to spatiotemporal attribute difference of thevisual object the self-occlusion avoidance problem can beclassified into static self-occlusion avoidance and dynamicself-occlusion avoidance Static self-occlusion avoidancerefers to the fact that when the visual object is staticthe self-occlusion phenomenon is avoided by moving thecamera to a next best view for observing the self-occlusionregion detected in an image of the visual object Dynamic
self-occlusion avoidance refers to the fact that when the visualobject is moving the self-occlusion phenomenon is avoidedby moving the camera step by step to the best view forobserving the self-occlusion region detected in one framefrom depth image sequence of the moving visual object Bythe concept above we can see themain difference between thetwo cases is that the result of static self-occlusion avoidanceis a next best view under self-occlusion and the cameracan accomplish the task of self-occlusion avoidance in thecalculated view which can be regarded as ldquoone-steprdquo typeWhile the result of dynamic self-occlusion avoidance is aseries of next views the camera needs multiple motions toavoid self-occlusion which can be regarded as ldquostep-by-steprdquotype
Compared with static self-occlusion avoidance thevisual object keeps moving in the process of dynamicself-occlusion avoidance which makes it impossible toaccomplish the dynamic self-occlusion avoidance throughmoving the camera to the calculated next best viewdirectly In order to accomplish the dynamic self-occlusion
Hindawi Publishing CorporationMathematical Problems in EngineeringVolume 2016 Article ID 4783794 11 pageshttpdxdoiorg10115520164783794
2 Mathematical Problems in Engineering
avoidance besides calculating the next views the motionestimation about the moving visual object is necessaryThen the camera not only does synchronous motionwith the visual object according to the motion estima-tion result but also executes self-occlusion avoidance func-tion simultaneously In this way the camera moves stepby step until accomplishing the dynamic self-occlusionavoidance
Before self-occlusion avoidance problem emerging thesimilar problemcalled next best viewhas beenwidely studiedCurrently scholars have gained some achievements on thenext best view Connolly [1] as one of the earlier scholarsstudying the next best view used partial octree model todescribe visual object and made different marks to the nodesof different observation situations so as to determine thenext best view By discretizing a fixed surface Pito [2]determined the next best view from plenty of discretizationcandidate views Based on the depth data Whaite and Ferrie[3] constructed parameter similarity model of object andplanned the next best view of the camera according to thedifference between the depth data and the current fittedmodel Li and Liu [4] proposed a method which used B-spline to construct the model of object and determined thenext best view by calculating information gain Trummeret al [5] proposed a next best view method by combiningon-line theory to optimize the 3D reconstruction precisionof any object Unfortunately because of never consideringocclusion factor in thesemethods themore serious the occlu-sion is the more the accuracy of these methods would beaffected
Because ubiquitous occlusion would affect the result ofthe next best view scholars further proposed the next bestview methods taking occlusion into account Banta et al[6] proposed a combination method to determine the nextbest view under occlusion Based on the integration of activeand passive vision Fang and He [7] determined the nextbest view through the shape of the shadow region andthe concept of limit visible surface Combining layered raytracing and octree Vasquez-Gomez et al [8] constructedthe object model and generated candidate views based onsorting of the utility function to determine the next bestview Potthast and Sukhatme [9] determined the next bestview through the difference of information entropies underocclusion Wu et al [10] determined the next best view byusing the algorithm of layered contour fitting (LCF) basedon density Although these methods consider the factor ofocclusion (mutual occlusion or self-occlusion this paperpays attention to self-occlusion) there are limitations in thecamera position [6 7] specific equipment [8 9] a prioriknowledge [7 10] and so forth And besides the self-occlusion region is not modeled properly in these next bestview methods Moreover there exist some differences (suchas problem description and solving cue) between the nextbest view and self-occlusion avoidance so the abovemethodshave some reference significance but cannot be taken as thefinal solution to the problem of self-occlusion avoidanceMost importantly all of the above methods aim at the studyof static visual object and they are not suitable for the
moving visual object That is to say all of the above methodscannot be the solution to dynamic self-occlusion avoidanceHowever in many scientific research fields such as real-time 3D reconstruction object tracking autonomous navi-gation scene recognition of mobile robot robot autonomousoperation and dynamic scene rendering the self-occlusionof moving visual object is a universal phenomenon Mean-while the visual system would be invalid even wrongif it cannot effectively detect and avoid self-occlusion ofmoving visual object thus the visual system would loseits value which makes how to avoid the self-occlusionof moving visual object become an inevitable and urgentproblem
According to investigation result there is no relatedresearch on dynamic self-occlusion avoidance of movingvisual object that is to say the research on dynamic self-occlusion avoidance is still in initial stage at present Mean-while considering the objective fact that the 3D informationof a scene can be better obtained from the depth imagethan the intensity image an approach for avoiding the self-occlusion of the rigid motion object is proposed in this paperbased on the depth image In the process of designing themethod wemainly solve the following three problems firstlyhow to design the solution to the dynamic self-occlusionavoidance problem secondly how to estimate the motionof visual object based on depth image thirdly how to mea-sure the effect of dynamic self-occlusion avoidance methodAiming at the first problem the dynamic self-occlusionavoidance is posed as an optimization problem and motionestimation is merged into the optimization process of bestview model to solve the problem of dynamic self-occlusionavoidance Aiming at the second problem the two Gaussiancurvature feature matrices are matched by SIFT algorithmand the efficient singular value decomposition (SVD) is usedto estimate themotion of visual object For the third problemldquoeffective avoidance raterdquo is proposed to measure the per-formance of the dynamic self-occlusion avoidance methodThe rest of the paper is organized as follows Section 2 isthe method overview Section 3 describes the dynamic self-occlusion avoidancemethod based on depth image Section 4presents our experimental results and Section 5 concludes thepaper
2 Method Overview
21 The Analysis of Dynamic Self-Occlusion AvoidanceThe problem of dynamic self-occlusion avoidance can bedescribed on the premise that the self-occlusion region inan image of visual object image sequence is taken as theresearch object the reasonable next view sequence is plannedfor observing the self-occlusion region bymoving the camerato achieve the goal that the most information about the self-occlusion region can be obtained from the camera final viewBecause the visual object is moving in the process of self-occlusion avoidance the next view of the camera shouldbe adjusted dynamically to observe the vested self-occlusionregion
Mathematical Problems in Engineering 3
Ideal object modelCamera
(a) (b) (c)
Figure 1 The sketch map of camera observing ideal object model(a) Observing effect in initial view (b) Observing effect duringavoiding self-occlusion (c) Observing effect at the end of avoidingself-occlusion
Figure 1 shows the sketch map of camera observing idealobject model Figure 1(a) is the observing effect of camerain initial view and the shadow region is the self-occlusionregion to be avoided In the case of object moving if camerawants to obtain the information of the self-occlusion regionit must do the synchronousmotion with the visual object andmove to the best view for observing self-occlusion region atthe same time Figure 1(b) is the observing effect when thecamera is avoiding self-occlusion Figure 1(c) is the observingeffect when the camera arrives at the final view in whichthe camera can obtain the maximum information of self-occlusion region so the process of dynamic self-occlusionavoidance is accomplished when the camera arrives at thisview
22 The Overall Idea of Dynamic Self-Occlusion AvoidanceBased on the analysis of dynamic self-occlusion avoidanceabove an approach to dynamic self-occlusion avoidance isproposed The overall idea of the approach is as followsFirstly two adjacent depth images of the moving objectare acquired and each pixelrsquos 3D coordinates (all pixelsrsquo 3Dcoordinates are in the same world coordinate system) intwo depth images are calculated by utilizing antiprojectiontransformation and the self-occlusion cue in the seconddepth image is detected On this basis the best view modelis constructed according to the self-occlusion informationin the second depth image Secondly according to the 3Dcoordinates calculated above the Gaussian curvature featurematrices corresponding to the two adjacent depth imagesare calculated by using the pixelsrsquo 3D coordinates Then theGaussian curvature feature matrices corresponding to thetwo adjacent depth images are matched by SIFT algorithmand the motion equation is estimated by using the 3Dcoordinates of the matched points Finally combining thebest view model and the estimated motion equation ofthe visual object a series of next views of the camera areplanned to accomplish dynamic self-occlusion avoidanceprocess
3 The Approach to Dynamic Self-OcclusionAvoidance Based on Depth Image
31 Constructing the Self-Occlusion Region to BeAvoided and the Best View Model
311 Constructing the Self-Occlusion Region to Be AvoidedIn order to solve the problem of dynamic self-occlusionavoidance it is necessary to first construct the self-occlusionregion to be avoided As we know the information of self-occlusion region in current view is unknown that is to saythe geometry information of self-occlusion region could notbe obtained directly according to the depth image acquiredin current view so the specific modeling method should beadopted to describe the self-occlusion region approximatelyFigure 2 shows the local self-occlusion region of visualobject and its approximate description and in Figure 2(a)the red boundary is self-occlusion boundary and the blueboundary is the nether adjacent boundary correspondingto the red one The method in literature [11] is used todetect the self-occlusion boundary in depth image and all thepoints on self-occlusion boundary are organized to form self-occlusion boundary point set119874 As shown in Figure 2(b) theregion between self-occlusion boundary and nether adjacentboundary in 3D space is the unknown self-occlusion regionFigure 2(c) shows the construction of self-occlusion regionmodel by utilizing quadrilateral subdivision on unknownself-occlusion region
The concrete steps of constructing self-occlusion regionmodel are as follows Firstly take out the 119894th self-occlusionboundary point 119900
119894from self-occlusion boundary point set 119874
in turn Mark the unmarked nonocclusion boundary pointcorresponding to themaximum of depth differences betweenthe self-occlusion boundary point 119900
119894and its eight neighbors
as 1199001015840119894 and add 119900
1015840
119894into nether adjacent boundary point set
1198741015840 Secondly use the neighboring self-occlusion boundary
points 119900119895 119900119895+1
in 119874 and their corresponding nether adjacentpoints 1199001015840
119895 1199001015840119895+1
in 1198741015840 to form quadrilateral patch
119895 as shown
in Figure 2(c) 119895 is the integer from 1 to 119873 and the com-bination of all 119873 quadrilaterals is approximate descriptionof self-occlusion region Thirdly combining depth imageand camera parameters use antiprojection transformationto calculate each pixelrsquos 3D coordinates At last use the 3Dcoordinates of quadrilateral patch
119895rsquos four vertices to calculate
its normal and area thus the modeling process of self-occlusion region can be accomplished
On the premise of not considering the normal directionthe formula for calculating the normal k1015840patch119895 of quadrilateralpatch119895can be defined as
k1015840patch119895 = u119895times w119895
or k1015840patch119895 = w119895times u119895
(1)
where u119895is the vector from point 119900
119895to the midpoint of 1199001015840
119895
and 1199001015840
119895+1in 3D space and w
119895is the vector from point 119900
119895to
point 119900119895+1
in 3D space Supposing the initial camera view is
4 Mathematical Problems in Engineering
oj
oj+1
o998400
j
o998400
j+1
patchj
Self-occlusion regionmodelUnknown self-
occlusion regionSelf-occlusion boundary
Upper surface
Nether surface
(a) (b) (c)
Modelling
Figure 2 The local self-occlusion region and its approximate description for visual object (a) The depth image of visual object (b) Spatialstructure of local self-occlusion region (c) Approximate description of self-occlusion region
(xbegin kbegin) in order to ensure that the direction of calcu-lated normal points outside in the process of self-occlusionregion modeling the calculation formula of patch
119895rsquos normal
kpatch119895 is defined as
kpatch119895 = sign (minusk1015840patch119895 sdot kbegin) sdot k1015840
patch119895 (2)
where sign is the sign function it means the function valueis 1 when minusk1015840patch119895 sdot kbegin ge 0 and the function value is minus1when minusk1015840patch119895 sdot kbegin lt 0 The area 119878
119895of quadrilateral patch
119895
is defined as
119878119895=100381710038171003817100381710038171003817kpatch119895
1003817100381710038171003817100381710038172 (3)
In summary the set 119875 consisting of all quadrilateralspatch119895is the modeling result of self-occlusion region so the
self-occlusion information to be avoided is described as
119875 = (119900119895 119900119895+1
1199001015840
119895 1199001015840
119895+1 vpatch119895 119878119895) | 119900
119895 119900119895+1
isin 119874 1199001015840
119895 1199001015840
119895+1isin 1198741015840
(4)
312 Constructing the Best ViewModel The best view modelshould be constructed after obtaining the self-occlusioninformation to be avoided This paper uses the area ofvisible self-occlusion region on next view to construct theobjective function and meanwhile uses the size equality ofreal object corresponding to pixels in different depth imagesas constraint condition So the best view model is defined as
argmaxxBV
119873
sum
119895=1
119878119895cos (120579
119895(xBV))
st 120579119895(xBV) =
120579119895(xBV) 120579
119895(xBV) lt
120587
2
120587
2120579119895(xBV) ge
120587
2
1003817100381710038171003817xmid minus xBV10038171003817100381710038172 =
10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172
(5)
where xBV and xmid minus xBV are respectively the position anddirection of camera for observing the vested self-occlusionregion at maximum level119873 is the number of quadrilaterals120579119895(xBV) is the angle between the normal of 119895th quadrilateral
and the vector from the midpoint of 119895th quadrilateral tocamera xmid is the center of gravity of all visible points ininitial camera view (xbegin kbegin)
32 Estimating theMotion Equation of the Visual Object Basedon Two Adjacent Depth Images Different from the static self-occlusion avoidance the dynamic self-occlusion avoidance iscarried out on the premise that the visual object is movingwhichmakes themotion estimation of the visual object in 3Dspace become a necessary step in the process of dynamic self-occlusion avoidance Because the existing motion estimationmethods based on depth image all estimate the motion ofthe visual object in 2D images this paper designs a feasiblescheme based on the depth image to estimate the motion ofthe visual object in 3D space
321 Calculating the Gaussian Curvature Feature Matricesof the Two Adjacent Depth Images Considering that rigidbody has rotation and translation invariant characteristicsthe surface shape of visual object remains invariant inthe motion process Because the curvature feature reflectsbending degree of the object surface and Gaussian curvatureonly depends on Riemannian metric of the surface which isthe intrinsic invariant of the surface (ie Gaussian curvaturekeeps invariant in rotation and translation) the motion ofvisual object is estimated based onGaussian curvature featurematrices corresponding to the two adjacent depth images
In 3D Euclidean space 1198961and 119896
2are two principal
curvatures of a point on the differentiable surface then theGaussian curvature is defined as the product of two principalcurvatures namely 119896 = 119896
11198962 The method of calculating
the Gaussian curvature in literature [12] is implemented byoptimizing convolution calculation and it is high-efficientand suitable for different scales of curvature value so the
Mathematical Problems in Engineering 5
method in literature [12] is used for calculating the Gaussiancurvature feature matrix of depth image
322 Estimating the Motion Equation of Visual Object Basedon the Matched Feature Points Analysis shows that thematched points can provide reliable information for themotion estimation if the Gaussian curvature feature matricescorresponding to two adjacent depth images of the visualobject can be matched SIFT algorithm in literature [13]is more efficient and it is not only invariant to imagetranslation rotation and scaling but also robust to thechange of visual angle affine transformation and noise sothe SIFT algorithm is adopted in this paper for matching thefeature points
SIFT algorithm mainly consists of two steps the gener-ation of SIFT features and the matching of feature vectorsIn order to generate the SIFT features firstly the scale spaceshould be established Because the Difference of Gaussiansscale-space (DoG scale-space) has good stability in detectionof key points the DoG scale-space is used for detecting thekey points in this paper Secondly the key points which areinstable and sensitive to noise are filtered Then the directionof each feature point is calculated Finally the SIFT featurevector of each key point is generated In the process ofgenerating SIFT feature vector 16 seed points are selectedfor each key point and each seed point has 8 orientationsTherefore the 16 times 8 = 128 dimensions of the SIFT featurevector can be obtained for each key point After obtainingthe SIFT feature vectors of the Gaussian curvature featurematrices corresponding to the two adjacent depth images thetwoGaussian curvature featurematrices can bematched andthe similarity of the vectors ismeasured byEuclidean distancein the process of matching
After the key points of the Gaussian curvature featurematrices corresponding to the two adjacent depth images arematched the motion equation of visual object can be esti-mated by using the pixelsrsquo 3D coordinates which are obtainedby antiprojection transformation The motion equation isdefined as
x1015840119897= Rx119897+ T (6)
where x119897and x1015840119897are respectively the 3D coordinates of points
before and after visual objectmotionR is the unit orthogonalrotation matrix and T is the translation vector then theresult of motion estimation is the solution of minimizing theobjective function 119891(RT) The objective function 119891(RT) isdefined as
119891 (RT) =119872
sum
119897=1
10038171003817100381710038171003817x1015840119897minus (Rx
119897+ T)10038171003817100381710038171003817
2
(7)
where119872 is the number of key points which are matched Inorder to solve the R and T in formula (7) the matrix H isdefined as
H =
119872
sum
119897=1
x1015840clx119879
cl (8)
where x1015840cl = x1015840119897minus (1119872)sum
119872
119897=1x1015840119897 xcl = x
119897minus (1119872)sum
119872
119897=1x119897
By carrying out singular value decomposition on H H canbe expressed as H=UΛV119879 then according to U and V R informula (7) can be deduced as
R = VU119879 (9)
After obtaining R substitute R into formula (10) tocalculate T namely
T = x1015840119897minus Rx119897 (10)
After obtaining R and T the motion equation shown informula (6) can be obtained and themotion estimation of thevisual object is accomplished
33 Planning the Next View Based on Motion Estimationand Best View Model After the motion equation of visualobject is estimated the next view of the camera can beplanned Analysis can be known if substituting the cameraobservation position into the motion equation of visualobject the camera can do the synchronous motion with thevisual object In addition on the basis of the synchronousmotion with the visual object if the camera motion offsetfor avoiding self-occlusion is introduced the whole processof dynamic self-occlusion avoidance can be achieved Byanalyzing the best view model in formula (5) we can knowit is the camera best view that can make formula (11) achievemaximum value Formula (11) is defined as
119891 (x) =119873
sum
119895=1
119878119895cos (120579
119895 (x)) (11)
where x is the camera observation position Since formula (11)is a nonconvex model it is difficult to directly calculate itsglobal optimal solution Meanwhile because the problem ofdynamic self-occlusion avoidance needs to calculate a seriesof next views and the gradient information in formula (11) canprovide basis for determining the cameramotion offset of theself-occlusion avoidance process in the process of dynamicself-occlusion avoidance the gradient descent idea is usedin this paper for defining the formula of camera observationposition x1015840
119873119881as
x1015840119873119881
= Rx119881+ T + 120575nabla119891 (x
119881) (12)
where x119881is the current observation position of the camera
and 120575 is the offset coefficient Based on the solution to formula(12) the formula of next observation direction k
119873119881is defined
as
k119873119881
= xmid minus x1015840119873119881
(13)
According to the constraint condition in formula (5) wecan search one point x
119873119881along the opposite direction of k
119873119881
and make x119873119881
meet1003817100381710038171003817xmid minus x
119873119881
10038171003817100381710038172 =10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172 (14)
By now the obtained view (x119873119881
v119873119881
) can be regardedas the new current view (x
119881 v119881) After acquiring the depth
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
2 Mathematical Problems in Engineering
avoidance besides calculating the next views the motionestimation about the moving visual object is necessaryThen the camera not only does synchronous motionwith the visual object according to the motion estima-tion result but also executes self-occlusion avoidance func-tion simultaneously In this way the camera moves stepby step until accomplishing the dynamic self-occlusionavoidance
Before self-occlusion avoidance problem emerging thesimilar problemcalled next best viewhas beenwidely studiedCurrently scholars have gained some achievements on thenext best view Connolly [1] as one of the earlier scholarsstudying the next best view used partial octree model todescribe visual object and made different marks to the nodesof different observation situations so as to determine thenext best view By discretizing a fixed surface Pito [2]determined the next best view from plenty of discretizationcandidate views Based on the depth data Whaite and Ferrie[3] constructed parameter similarity model of object andplanned the next best view of the camera according to thedifference between the depth data and the current fittedmodel Li and Liu [4] proposed a method which used B-spline to construct the model of object and determined thenext best view by calculating information gain Trummeret al [5] proposed a next best view method by combiningon-line theory to optimize the 3D reconstruction precisionof any object Unfortunately because of never consideringocclusion factor in thesemethods themore serious the occlu-sion is the more the accuracy of these methods would beaffected
Because ubiquitous occlusion would affect the result ofthe next best view scholars further proposed the next bestview methods taking occlusion into account Banta et al[6] proposed a combination method to determine the nextbest view under occlusion Based on the integration of activeand passive vision Fang and He [7] determined the nextbest view through the shape of the shadow region andthe concept of limit visible surface Combining layered raytracing and octree Vasquez-Gomez et al [8] constructedthe object model and generated candidate views based onsorting of the utility function to determine the next bestview Potthast and Sukhatme [9] determined the next bestview through the difference of information entropies underocclusion Wu et al [10] determined the next best view byusing the algorithm of layered contour fitting (LCF) basedon density Although these methods consider the factor ofocclusion (mutual occlusion or self-occlusion this paperpays attention to self-occlusion) there are limitations in thecamera position [6 7] specific equipment [8 9] a prioriknowledge [7 10] and so forth And besides the self-occlusion region is not modeled properly in these next bestview methods Moreover there exist some differences (suchas problem description and solving cue) between the nextbest view and self-occlusion avoidance so the abovemethodshave some reference significance but cannot be taken as thefinal solution to the problem of self-occlusion avoidanceMost importantly all of the above methods aim at the studyof static visual object and they are not suitable for the
moving visual object That is to say all of the above methodscannot be the solution to dynamic self-occlusion avoidanceHowever in many scientific research fields such as real-time 3D reconstruction object tracking autonomous navi-gation scene recognition of mobile robot robot autonomousoperation and dynamic scene rendering the self-occlusionof moving visual object is a universal phenomenon Mean-while the visual system would be invalid even wrongif it cannot effectively detect and avoid self-occlusion ofmoving visual object thus the visual system would loseits value which makes how to avoid the self-occlusionof moving visual object become an inevitable and urgentproblem
According to investigation result there is no relatedresearch on dynamic self-occlusion avoidance of movingvisual object that is to say the research on dynamic self-occlusion avoidance is still in initial stage at present Mean-while considering the objective fact that the 3D informationof a scene can be better obtained from the depth imagethan the intensity image an approach for avoiding the self-occlusion of the rigid motion object is proposed in this paperbased on the depth image In the process of designing themethod wemainly solve the following three problems firstlyhow to design the solution to the dynamic self-occlusionavoidance problem secondly how to estimate the motionof visual object based on depth image thirdly how to mea-sure the effect of dynamic self-occlusion avoidance methodAiming at the first problem the dynamic self-occlusionavoidance is posed as an optimization problem and motionestimation is merged into the optimization process of bestview model to solve the problem of dynamic self-occlusionavoidance Aiming at the second problem the two Gaussiancurvature feature matrices are matched by SIFT algorithmand the efficient singular value decomposition (SVD) is usedto estimate themotion of visual object For the third problemldquoeffective avoidance raterdquo is proposed to measure the per-formance of the dynamic self-occlusion avoidance methodThe rest of the paper is organized as follows Section 2 isthe method overview Section 3 describes the dynamic self-occlusion avoidancemethod based on depth image Section 4presents our experimental results and Section 5 concludes thepaper
2 Method Overview
21 The Analysis of Dynamic Self-Occlusion AvoidanceThe problem of dynamic self-occlusion avoidance can bedescribed on the premise that the self-occlusion region inan image of visual object image sequence is taken as theresearch object the reasonable next view sequence is plannedfor observing the self-occlusion region bymoving the camerato achieve the goal that the most information about the self-occlusion region can be obtained from the camera final viewBecause the visual object is moving in the process of self-occlusion avoidance the next view of the camera shouldbe adjusted dynamically to observe the vested self-occlusionregion
Mathematical Problems in Engineering 3
Ideal object modelCamera
(a) (b) (c)
Figure 1 The sketch map of camera observing ideal object model(a) Observing effect in initial view (b) Observing effect duringavoiding self-occlusion (c) Observing effect at the end of avoidingself-occlusion
Figure 1 shows the sketch map of camera observing idealobject model Figure 1(a) is the observing effect of camerain initial view and the shadow region is the self-occlusionregion to be avoided In the case of object moving if camerawants to obtain the information of the self-occlusion regionit must do the synchronousmotion with the visual object andmove to the best view for observing self-occlusion region atthe same time Figure 1(b) is the observing effect when thecamera is avoiding self-occlusion Figure 1(c) is the observingeffect when the camera arrives at the final view in whichthe camera can obtain the maximum information of self-occlusion region so the process of dynamic self-occlusionavoidance is accomplished when the camera arrives at thisview
22 The Overall Idea of Dynamic Self-Occlusion AvoidanceBased on the analysis of dynamic self-occlusion avoidanceabove an approach to dynamic self-occlusion avoidance isproposed The overall idea of the approach is as followsFirstly two adjacent depth images of the moving objectare acquired and each pixelrsquos 3D coordinates (all pixelsrsquo 3Dcoordinates are in the same world coordinate system) intwo depth images are calculated by utilizing antiprojectiontransformation and the self-occlusion cue in the seconddepth image is detected On this basis the best view modelis constructed according to the self-occlusion informationin the second depth image Secondly according to the 3Dcoordinates calculated above the Gaussian curvature featurematrices corresponding to the two adjacent depth imagesare calculated by using the pixelsrsquo 3D coordinates Then theGaussian curvature feature matrices corresponding to thetwo adjacent depth images are matched by SIFT algorithmand the motion equation is estimated by using the 3Dcoordinates of the matched points Finally combining thebest view model and the estimated motion equation ofthe visual object a series of next views of the camera areplanned to accomplish dynamic self-occlusion avoidanceprocess
3 The Approach to Dynamic Self-OcclusionAvoidance Based on Depth Image
31 Constructing the Self-Occlusion Region to BeAvoided and the Best View Model
311 Constructing the Self-Occlusion Region to Be AvoidedIn order to solve the problem of dynamic self-occlusionavoidance it is necessary to first construct the self-occlusionregion to be avoided As we know the information of self-occlusion region in current view is unknown that is to saythe geometry information of self-occlusion region could notbe obtained directly according to the depth image acquiredin current view so the specific modeling method should beadopted to describe the self-occlusion region approximatelyFigure 2 shows the local self-occlusion region of visualobject and its approximate description and in Figure 2(a)the red boundary is self-occlusion boundary and the blueboundary is the nether adjacent boundary correspondingto the red one The method in literature [11] is used todetect the self-occlusion boundary in depth image and all thepoints on self-occlusion boundary are organized to form self-occlusion boundary point set119874 As shown in Figure 2(b) theregion between self-occlusion boundary and nether adjacentboundary in 3D space is the unknown self-occlusion regionFigure 2(c) shows the construction of self-occlusion regionmodel by utilizing quadrilateral subdivision on unknownself-occlusion region
The concrete steps of constructing self-occlusion regionmodel are as follows Firstly take out the 119894th self-occlusionboundary point 119900
119894from self-occlusion boundary point set 119874
in turn Mark the unmarked nonocclusion boundary pointcorresponding to themaximum of depth differences betweenthe self-occlusion boundary point 119900
119894and its eight neighbors
as 1199001015840119894 and add 119900
1015840
119894into nether adjacent boundary point set
1198741015840 Secondly use the neighboring self-occlusion boundary
points 119900119895 119900119895+1
in 119874 and their corresponding nether adjacentpoints 1199001015840
119895 1199001015840119895+1
in 1198741015840 to form quadrilateral patch
119895 as shown
in Figure 2(c) 119895 is the integer from 1 to 119873 and the com-bination of all 119873 quadrilaterals is approximate descriptionof self-occlusion region Thirdly combining depth imageand camera parameters use antiprojection transformationto calculate each pixelrsquos 3D coordinates At last use the 3Dcoordinates of quadrilateral patch
119895rsquos four vertices to calculate
its normal and area thus the modeling process of self-occlusion region can be accomplished
On the premise of not considering the normal directionthe formula for calculating the normal k1015840patch119895 of quadrilateralpatch119895can be defined as
k1015840patch119895 = u119895times w119895
or k1015840patch119895 = w119895times u119895
(1)
where u119895is the vector from point 119900
119895to the midpoint of 1199001015840
119895
and 1199001015840
119895+1in 3D space and w
119895is the vector from point 119900
119895to
point 119900119895+1
in 3D space Supposing the initial camera view is
4 Mathematical Problems in Engineering
oj
oj+1
o998400
j
o998400
j+1
patchj
Self-occlusion regionmodelUnknown self-
occlusion regionSelf-occlusion boundary
Upper surface
Nether surface
(a) (b) (c)
Modelling
Figure 2 The local self-occlusion region and its approximate description for visual object (a) The depth image of visual object (b) Spatialstructure of local self-occlusion region (c) Approximate description of self-occlusion region
(xbegin kbegin) in order to ensure that the direction of calcu-lated normal points outside in the process of self-occlusionregion modeling the calculation formula of patch
119895rsquos normal
kpatch119895 is defined as
kpatch119895 = sign (minusk1015840patch119895 sdot kbegin) sdot k1015840
patch119895 (2)
where sign is the sign function it means the function valueis 1 when minusk1015840patch119895 sdot kbegin ge 0 and the function value is minus1when minusk1015840patch119895 sdot kbegin lt 0 The area 119878
119895of quadrilateral patch
119895
is defined as
119878119895=100381710038171003817100381710038171003817kpatch119895
1003817100381710038171003817100381710038172 (3)
In summary the set 119875 consisting of all quadrilateralspatch119895is the modeling result of self-occlusion region so the
self-occlusion information to be avoided is described as
119875 = (119900119895 119900119895+1
1199001015840
119895 1199001015840
119895+1 vpatch119895 119878119895) | 119900
119895 119900119895+1
isin 119874 1199001015840
119895 1199001015840
119895+1isin 1198741015840
(4)
312 Constructing the Best ViewModel The best view modelshould be constructed after obtaining the self-occlusioninformation to be avoided This paper uses the area ofvisible self-occlusion region on next view to construct theobjective function and meanwhile uses the size equality ofreal object corresponding to pixels in different depth imagesas constraint condition So the best view model is defined as
argmaxxBV
119873
sum
119895=1
119878119895cos (120579
119895(xBV))
st 120579119895(xBV) =
120579119895(xBV) 120579
119895(xBV) lt
120587
2
120587
2120579119895(xBV) ge
120587
2
1003817100381710038171003817xmid minus xBV10038171003817100381710038172 =
10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172
(5)
where xBV and xmid minus xBV are respectively the position anddirection of camera for observing the vested self-occlusionregion at maximum level119873 is the number of quadrilaterals120579119895(xBV) is the angle between the normal of 119895th quadrilateral
and the vector from the midpoint of 119895th quadrilateral tocamera xmid is the center of gravity of all visible points ininitial camera view (xbegin kbegin)
32 Estimating theMotion Equation of the Visual Object Basedon Two Adjacent Depth Images Different from the static self-occlusion avoidance the dynamic self-occlusion avoidance iscarried out on the premise that the visual object is movingwhichmakes themotion estimation of the visual object in 3Dspace become a necessary step in the process of dynamic self-occlusion avoidance Because the existing motion estimationmethods based on depth image all estimate the motion ofthe visual object in 2D images this paper designs a feasiblescheme based on the depth image to estimate the motion ofthe visual object in 3D space
321 Calculating the Gaussian Curvature Feature Matricesof the Two Adjacent Depth Images Considering that rigidbody has rotation and translation invariant characteristicsthe surface shape of visual object remains invariant inthe motion process Because the curvature feature reflectsbending degree of the object surface and Gaussian curvatureonly depends on Riemannian metric of the surface which isthe intrinsic invariant of the surface (ie Gaussian curvaturekeeps invariant in rotation and translation) the motion ofvisual object is estimated based onGaussian curvature featurematrices corresponding to the two adjacent depth images
In 3D Euclidean space 1198961and 119896
2are two principal
curvatures of a point on the differentiable surface then theGaussian curvature is defined as the product of two principalcurvatures namely 119896 = 119896
11198962 The method of calculating
the Gaussian curvature in literature [12] is implemented byoptimizing convolution calculation and it is high-efficientand suitable for different scales of curvature value so the
Mathematical Problems in Engineering 5
method in literature [12] is used for calculating the Gaussiancurvature feature matrix of depth image
322 Estimating the Motion Equation of Visual Object Basedon the Matched Feature Points Analysis shows that thematched points can provide reliable information for themotion estimation if the Gaussian curvature feature matricescorresponding to two adjacent depth images of the visualobject can be matched SIFT algorithm in literature [13]is more efficient and it is not only invariant to imagetranslation rotation and scaling but also robust to thechange of visual angle affine transformation and noise sothe SIFT algorithm is adopted in this paper for matching thefeature points
SIFT algorithm mainly consists of two steps the gener-ation of SIFT features and the matching of feature vectorsIn order to generate the SIFT features firstly the scale spaceshould be established Because the Difference of Gaussiansscale-space (DoG scale-space) has good stability in detectionof key points the DoG scale-space is used for detecting thekey points in this paper Secondly the key points which areinstable and sensitive to noise are filtered Then the directionof each feature point is calculated Finally the SIFT featurevector of each key point is generated In the process ofgenerating SIFT feature vector 16 seed points are selectedfor each key point and each seed point has 8 orientationsTherefore the 16 times 8 = 128 dimensions of the SIFT featurevector can be obtained for each key point After obtainingthe SIFT feature vectors of the Gaussian curvature featurematrices corresponding to the two adjacent depth images thetwoGaussian curvature featurematrices can bematched andthe similarity of the vectors ismeasured byEuclidean distancein the process of matching
After the key points of the Gaussian curvature featurematrices corresponding to the two adjacent depth images arematched the motion equation of visual object can be esti-mated by using the pixelsrsquo 3D coordinates which are obtainedby antiprojection transformation The motion equation isdefined as
x1015840119897= Rx119897+ T (6)
where x119897and x1015840119897are respectively the 3D coordinates of points
before and after visual objectmotionR is the unit orthogonalrotation matrix and T is the translation vector then theresult of motion estimation is the solution of minimizing theobjective function 119891(RT) The objective function 119891(RT) isdefined as
119891 (RT) =119872
sum
119897=1
10038171003817100381710038171003817x1015840119897minus (Rx
119897+ T)10038171003817100381710038171003817
2
(7)
where119872 is the number of key points which are matched Inorder to solve the R and T in formula (7) the matrix H isdefined as
H =
119872
sum
119897=1
x1015840clx119879
cl (8)
where x1015840cl = x1015840119897minus (1119872)sum
119872
119897=1x1015840119897 xcl = x
119897minus (1119872)sum
119872
119897=1x119897
By carrying out singular value decomposition on H H canbe expressed as H=UΛV119879 then according to U and V R informula (7) can be deduced as
R = VU119879 (9)
After obtaining R substitute R into formula (10) tocalculate T namely
T = x1015840119897minus Rx119897 (10)
After obtaining R and T the motion equation shown informula (6) can be obtained and themotion estimation of thevisual object is accomplished
33 Planning the Next View Based on Motion Estimationand Best View Model After the motion equation of visualobject is estimated the next view of the camera can beplanned Analysis can be known if substituting the cameraobservation position into the motion equation of visualobject the camera can do the synchronous motion with thevisual object In addition on the basis of the synchronousmotion with the visual object if the camera motion offsetfor avoiding self-occlusion is introduced the whole processof dynamic self-occlusion avoidance can be achieved Byanalyzing the best view model in formula (5) we can knowit is the camera best view that can make formula (11) achievemaximum value Formula (11) is defined as
119891 (x) =119873
sum
119895=1
119878119895cos (120579
119895 (x)) (11)
where x is the camera observation position Since formula (11)is a nonconvex model it is difficult to directly calculate itsglobal optimal solution Meanwhile because the problem ofdynamic self-occlusion avoidance needs to calculate a seriesof next views and the gradient information in formula (11) canprovide basis for determining the cameramotion offset of theself-occlusion avoidance process in the process of dynamicself-occlusion avoidance the gradient descent idea is usedin this paper for defining the formula of camera observationposition x1015840
119873119881as
x1015840119873119881
= Rx119881+ T + 120575nabla119891 (x
119881) (12)
where x119881is the current observation position of the camera
and 120575 is the offset coefficient Based on the solution to formula(12) the formula of next observation direction k
119873119881is defined
as
k119873119881
= xmid minus x1015840119873119881
(13)
According to the constraint condition in formula (5) wecan search one point x
119873119881along the opposite direction of k
119873119881
and make x119873119881
meet1003817100381710038171003817xmid minus x
119873119881
10038171003817100381710038172 =10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172 (14)
By now the obtained view (x119873119881
v119873119881
) can be regardedas the new current view (x
119881 v119881) After acquiring the depth
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Mathematical Problems in Engineering 3
Ideal object modelCamera
(a) (b) (c)
Figure 1 The sketch map of camera observing ideal object model(a) Observing effect in initial view (b) Observing effect duringavoiding self-occlusion (c) Observing effect at the end of avoidingself-occlusion
Figure 1 shows the sketch map of camera observing idealobject model Figure 1(a) is the observing effect of camerain initial view and the shadow region is the self-occlusionregion to be avoided In the case of object moving if camerawants to obtain the information of the self-occlusion regionit must do the synchronousmotion with the visual object andmove to the best view for observing self-occlusion region atthe same time Figure 1(b) is the observing effect when thecamera is avoiding self-occlusion Figure 1(c) is the observingeffect when the camera arrives at the final view in whichthe camera can obtain the maximum information of self-occlusion region so the process of dynamic self-occlusionavoidance is accomplished when the camera arrives at thisview
22 The Overall Idea of Dynamic Self-Occlusion AvoidanceBased on the analysis of dynamic self-occlusion avoidanceabove an approach to dynamic self-occlusion avoidance isproposed The overall idea of the approach is as followsFirstly two adjacent depth images of the moving objectare acquired and each pixelrsquos 3D coordinates (all pixelsrsquo 3Dcoordinates are in the same world coordinate system) intwo depth images are calculated by utilizing antiprojectiontransformation and the self-occlusion cue in the seconddepth image is detected On this basis the best view modelis constructed according to the self-occlusion informationin the second depth image Secondly according to the 3Dcoordinates calculated above the Gaussian curvature featurematrices corresponding to the two adjacent depth imagesare calculated by using the pixelsrsquo 3D coordinates Then theGaussian curvature feature matrices corresponding to thetwo adjacent depth images are matched by SIFT algorithmand the motion equation is estimated by using the 3Dcoordinates of the matched points Finally combining thebest view model and the estimated motion equation ofthe visual object a series of next views of the camera areplanned to accomplish dynamic self-occlusion avoidanceprocess
3 The Approach to Dynamic Self-OcclusionAvoidance Based on Depth Image
31 Constructing the Self-Occlusion Region to BeAvoided and the Best View Model
311 Constructing the Self-Occlusion Region to Be AvoidedIn order to solve the problem of dynamic self-occlusionavoidance it is necessary to first construct the self-occlusionregion to be avoided As we know the information of self-occlusion region in current view is unknown that is to saythe geometry information of self-occlusion region could notbe obtained directly according to the depth image acquiredin current view so the specific modeling method should beadopted to describe the self-occlusion region approximatelyFigure 2 shows the local self-occlusion region of visualobject and its approximate description and in Figure 2(a)the red boundary is self-occlusion boundary and the blueboundary is the nether adjacent boundary correspondingto the red one The method in literature [11] is used todetect the self-occlusion boundary in depth image and all thepoints on self-occlusion boundary are organized to form self-occlusion boundary point set119874 As shown in Figure 2(b) theregion between self-occlusion boundary and nether adjacentboundary in 3D space is the unknown self-occlusion regionFigure 2(c) shows the construction of self-occlusion regionmodel by utilizing quadrilateral subdivision on unknownself-occlusion region
The concrete steps of constructing self-occlusion regionmodel are as follows Firstly take out the 119894th self-occlusionboundary point 119900
119894from self-occlusion boundary point set 119874
in turn Mark the unmarked nonocclusion boundary pointcorresponding to themaximum of depth differences betweenthe self-occlusion boundary point 119900
119894and its eight neighbors
as 1199001015840119894 and add 119900
1015840
119894into nether adjacent boundary point set
1198741015840 Secondly use the neighboring self-occlusion boundary
points 119900119895 119900119895+1
in 119874 and their corresponding nether adjacentpoints 1199001015840
119895 1199001015840119895+1
in 1198741015840 to form quadrilateral patch
119895 as shown
in Figure 2(c) 119895 is the integer from 1 to 119873 and the com-bination of all 119873 quadrilaterals is approximate descriptionof self-occlusion region Thirdly combining depth imageand camera parameters use antiprojection transformationto calculate each pixelrsquos 3D coordinates At last use the 3Dcoordinates of quadrilateral patch
119895rsquos four vertices to calculate
its normal and area thus the modeling process of self-occlusion region can be accomplished
On the premise of not considering the normal directionthe formula for calculating the normal k1015840patch119895 of quadrilateralpatch119895can be defined as
k1015840patch119895 = u119895times w119895
or k1015840patch119895 = w119895times u119895
(1)
where u119895is the vector from point 119900
119895to the midpoint of 1199001015840
119895
and 1199001015840
119895+1in 3D space and w
119895is the vector from point 119900
119895to
point 119900119895+1
in 3D space Supposing the initial camera view is
4 Mathematical Problems in Engineering
oj
oj+1
o998400
j
o998400
j+1
patchj
Self-occlusion regionmodelUnknown self-
occlusion regionSelf-occlusion boundary
Upper surface
Nether surface
(a) (b) (c)
Modelling
Figure 2 The local self-occlusion region and its approximate description for visual object (a) The depth image of visual object (b) Spatialstructure of local self-occlusion region (c) Approximate description of self-occlusion region
(xbegin kbegin) in order to ensure that the direction of calcu-lated normal points outside in the process of self-occlusionregion modeling the calculation formula of patch
119895rsquos normal
kpatch119895 is defined as
kpatch119895 = sign (minusk1015840patch119895 sdot kbegin) sdot k1015840
patch119895 (2)
where sign is the sign function it means the function valueis 1 when minusk1015840patch119895 sdot kbegin ge 0 and the function value is minus1when minusk1015840patch119895 sdot kbegin lt 0 The area 119878
119895of quadrilateral patch
119895
is defined as
119878119895=100381710038171003817100381710038171003817kpatch119895
1003817100381710038171003817100381710038172 (3)
In summary the set 119875 consisting of all quadrilateralspatch119895is the modeling result of self-occlusion region so the
self-occlusion information to be avoided is described as
119875 = (119900119895 119900119895+1
1199001015840
119895 1199001015840
119895+1 vpatch119895 119878119895) | 119900
119895 119900119895+1
isin 119874 1199001015840
119895 1199001015840
119895+1isin 1198741015840
(4)
312 Constructing the Best ViewModel The best view modelshould be constructed after obtaining the self-occlusioninformation to be avoided This paper uses the area ofvisible self-occlusion region on next view to construct theobjective function and meanwhile uses the size equality ofreal object corresponding to pixels in different depth imagesas constraint condition So the best view model is defined as
argmaxxBV
119873
sum
119895=1
119878119895cos (120579
119895(xBV))
st 120579119895(xBV) =
120579119895(xBV) 120579
119895(xBV) lt
120587
2
120587
2120579119895(xBV) ge
120587
2
1003817100381710038171003817xmid minus xBV10038171003817100381710038172 =
10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172
(5)
where xBV and xmid minus xBV are respectively the position anddirection of camera for observing the vested self-occlusionregion at maximum level119873 is the number of quadrilaterals120579119895(xBV) is the angle between the normal of 119895th quadrilateral
and the vector from the midpoint of 119895th quadrilateral tocamera xmid is the center of gravity of all visible points ininitial camera view (xbegin kbegin)
32 Estimating theMotion Equation of the Visual Object Basedon Two Adjacent Depth Images Different from the static self-occlusion avoidance the dynamic self-occlusion avoidance iscarried out on the premise that the visual object is movingwhichmakes themotion estimation of the visual object in 3Dspace become a necessary step in the process of dynamic self-occlusion avoidance Because the existing motion estimationmethods based on depth image all estimate the motion ofthe visual object in 2D images this paper designs a feasiblescheme based on the depth image to estimate the motion ofthe visual object in 3D space
321 Calculating the Gaussian Curvature Feature Matricesof the Two Adjacent Depth Images Considering that rigidbody has rotation and translation invariant characteristicsthe surface shape of visual object remains invariant inthe motion process Because the curvature feature reflectsbending degree of the object surface and Gaussian curvatureonly depends on Riemannian metric of the surface which isthe intrinsic invariant of the surface (ie Gaussian curvaturekeeps invariant in rotation and translation) the motion ofvisual object is estimated based onGaussian curvature featurematrices corresponding to the two adjacent depth images
In 3D Euclidean space 1198961and 119896
2are two principal
curvatures of a point on the differentiable surface then theGaussian curvature is defined as the product of two principalcurvatures namely 119896 = 119896
11198962 The method of calculating
the Gaussian curvature in literature [12] is implemented byoptimizing convolution calculation and it is high-efficientand suitable for different scales of curvature value so the
Mathematical Problems in Engineering 5
method in literature [12] is used for calculating the Gaussiancurvature feature matrix of depth image
322 Estimating the Motion Equation of Visual Object Basedon the Matched Feature Points Analysis shows that thematched points can provide reliable information for themotion estimation if the Gaussian curvature feature matricescorresponding to two adjacent depth images of the visualobject can be matched SIFT algorithm in literature [13]is more efficient and it is not only invariant to imagetranslation rotation and scaling but also robust to thechange of visual angle affine transformation and noise sothe SIFT algorithm is adopted in this paper for matching thefeature points
SIFT algorithm mainly consists of two steps the gener-ation of SIFT features and the matching of feature vectorsIn order to generate the SIFT features firstly the scale spaceshould be established Because the Difference of Gaussiansscale-space (DoG scale-space) has good stability in detectionof key points the DoG scale-space is used for detecting thekey points in this paper Secondly the key points which areinstable and sensitive to noise are filtered Then the directionof each feature point is calculated Finally the SIFT featurevector of each key point is generated In the process ofgenerating SIFT feature vector 16 seed points are selectedfor each key point and each seed point has 8 orientationsTherefore the 16 times 8 = 128 dimensions of the SIFT featurevector can be obtained for each key point After obtainingthe SIFT feature vectors of the Gaussian curvature featurematrices corresponding to the two adjacent depth images thetwoGaussian curvature featurematrices can bematched andthe similarity of the vectors ismeasured byEuclidean distancein the process of matching
After the key points of the Gaussian curvature featurematrices corresponding to the two adjacent depth images arematched the motion equation of visual object can be esti-mated by using the pixelsrsquo 3D coordinates which are obtainedby antiprojection transformation The motion equation isdefined as
x1015840119897= Rx119897+ T (6)
where x119897and x1015840119897are respectively the 3D coordinates of points
before and after visual objectmotionR is the unit orthogonalrotation matrix and T is the translation vector then theresult of motion estimation is the solution of minimizing theobjective function 119891(RT) The objective function 119891(RT) isdefined as
119891 (RT) =119872
sum
119897=1
10038171003817100381710038171003817x1015840119897minus (Rx
119897+ T)10038171003817100381710038171003817
2
(7)
where119872 is the number of key points which are matched Inorder to solve the R and T in formula (7) the matrix H isdefined as
H =
119872
sum
119897=1
x1015840clx119879
cl (8)
where x1015840cl = x1015840119897minus (1119872)sum
119872
119897=1x1015840119897 xcl = x
119897minus (1119872)sum
119872
119897=1x119897
By carrying out singular value decomposition on H H canbe expressed as H=UΛV119879 then according to U and V R informula (7) can be deduced as
R = VU119879 (9)
After obtaining R substitute R into formula (10) tocalculate T namely
T = x1015840119897minus Rx119897 (10)
After obtaining R and T the motion equation shown informula (6) can be obtained and themotion estimation of thevisual object is accomplished
33 Planning the Next View Based on Motion Estimationand Best View Model After the motion equation of visualobject is estimated the next view of the camera can beplanned Analysis can be known if substituting the cameraobservation position into the motion equation of visualobject the camera can do the synchronous motion with thevisual object In addition on the basis of the synchronousmotion with the visual object if the camera motion offsetfor avoiding self-occlusion is introduced the whole processof dynamic self-occlusion avoidance can be achieved Byanalyzing the best view model in formula (5) we can knowit is the camera best view that can make formula (11) achievemaximum value Formula (11) is defined as
119891 (x) =119873
sum
119895=1
119878119895cos (120579
119895 (x)) (11)
where x is the camera observation position Since formula (11)is a nonconvex model it is difficult to directly calculate itsglobal optimal solution Meanwhile because the problem ofdynamic self-occlusion avoidance needs to calculate a seriesof next views and the gradient information in formula (11) canprovide basis for determining the cameramotion offset of theself-occlusion avoidance process in the process of dynamicself-occlusion avoidance the gradient descent idea is usedin this paper for defining the formula of camera observationposition x1015840
119873119881as
x1015840119873119881
= Rx119881+ T + 120575nabla119891 (x
119881) (12)
where x119881is the current observation position of the camera
and 120575 is the offset coefficient Based on the solution to formula(12) the formula of next observation direction k
119873119881is defined
as
k119873119881
= xmid minus x1015840119873119881
(13)
According to the constraint condition in formula (5) wecan search one point x
119873119881along the opposite direction of k
119873119881
and make x119873119881
meet1003817100381710038171003817xmid minus x
119873119881
10038171003817100381710038172 =10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172 (14)
By now the obtained view (x119873119881
v119873119881
) can be regardedas the new current view (x
119881 v119881) After acquiring the depth
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
4 Mathematical Problems in Engineering
oj
oj+1
o998400
j
o998400
j+1
patchj
Self-occlusion regionmodelUnknown self-
occlusion regionSelf-occlusion boundary
Upper surface
Nether surface
(a) (b) (c)
Modelling
Figure 2 The local self-occlusion region and its approximate description for visual object (a) The depth image of visual object (b) Spatialstructure of local self-occlusion region (c) Approximate description of self-occlusion region
(xbegin kbegin) in order to ensure that the direction of calcu-lated normal points outside in the process of self-occlusionregion modeling the calculation formula of patch
119895rsquos normal
kpatch119895 is defined as
kpatch119895 = sign (minusk1015840patch119895 sdot kbegin) sdot k1015840
patch119895 (2)
where sign is the sign function it means the function valueis 1 when minusk1015840patch119895 sdot kbegin ge 0 and the function value is minus1when minusk1015840patch119895 sdot kbegin lt 0 The area 119878
119895of quadrilateral patch
119895
is defined as
119878119895=100381710038171003817100381710038171003817kpatch119895
1003817100381710038171003817100381710038172 (3)
In summary the set 119875 consisting of all quadrilateralspatch119895is the modeling result of self-occlusion region so the
self-occlusion information to be avoided is described as
119875 = (119900119895 119900119895+1
1199001015840
119895 1199001015840
119895+1 vpatch119895 119878119895) | 119900
119895 119900119895+1
isin 119874 1199001015840
119895 1199001015840
119895+1isin 1198741015840
(4)
312 Constructing the Best ViewModel The best view modelshould be constructed after obtaining the self-occlusioninformation to be avoided This paper uses the area ofvisible self-occlusion region on next view to construct theobjective function and meanwhile uses the size equality ofreal object corresponding to pixels in different depth imagesas constraint condition So the best view model is defined as
argmaxxBV
119873
sum
119895=1
119878119895cos (120579
119895(xBV))
st 120579119895(xBV) =
120579119895(xBV) 120579
119895(xBV) lt
120587
2
120587
2120579119895(xBV) ge
120587
2
1003817100381710038171003817xmid minus xBV10038171003817100381710038172 =
10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172
(5)
where xBV and xmid minus xBV are respectively the position anddirection of camera for observing the vested self-occlusionregion at maximum level119873 is the number of quadrilaterals120579119895(xBV) is the angle between the normal of 119895th quadrilateral
and the vector from the midpoint of 119895th quadrilateral tocamera xmid is the center of gravity of all visible points ininitial camera view (xbegin kbegin)
32 Estimating theMotion Equation of the Visual Object Basedon Two Adjacent Depth Images Different from the static self-occlusion avoidance the dynamic self-occlusion avoidance iscarried out on the premise that the visual object is movingwhichmakes themotion estimation of the visual object in 3Dspace become a necessary step in the process of dynamic self-occlusion avoidance Because the existing motion estimationmethods based on depth image all estimate the motion ofthe visual object in 2D images this paper designs a feasiblescheme based on the depth image to estimate the motion ofthe visual object in 3D space
321 Calculating the Gaussian Curvature Feature Matricesof the Two Adjacent Depth Images Considering that rigidbody has rotation and translation invariant characteristicsthe surface shape of visual object remains invariant inthe motion process Because the curvature feature reflectsbending degree of the object surface and Gaussian curvatureonly depends on Riemannian metric of the surface which isthe intrinsic invariant of the surface (ie Gaussian curvaturekeeps invariant in rotation and translation) the motion ofvisual object is estimated based onGaussian curvature featurematrices corresponding to the two adjacent depth images
In 3D Euclidean space 1198961and 119896
2are two principal
curvatures of a point on the differentiable surface then theGaussian curvature is defined as the product of two principalcurvatures namely 119896 = 119896
11198962 The method of calculating
the Gaussian curvature in literature [12] is implemented byoptimizing convolution calculation and it is high-efficientand suitable for different scales of curvature value so the
Mathematical Problems in Engineering 5
method in literature [12] is used for calculating the Gaussiancurvature feature matrix of depth image
322 Estimating the Motion Equation of Visual Object Basedon the Matched Feature Points Analysis shows that thematched points can provide reliable information for themotion estimation if the Gaussian curvature feature matricescorresponding to two adjacent depth images of the visualobject can be matched SIFT algorithm in literature [13]is more efficient and it is not only invariant to imagetranslation rotation and scaling but also robust to thechange of visual angle affine transformation and noise sothe SIFT algorithm is adopted in this paper for matching thefeature points
SIFT algorithm mainly consists of two steps the gener-ation of SIFT features and the matching of feature vectorsIn order to generate the SIFT features firstly the scale spaceshould be established Because the Difference of Gaussiansscale-space (DoG scale-space) has good stability in detectionof key points the DoG scale-space is used for detecting thekey points in this paper Secondly the key points which areinstable and sensitive to noise are filtered Then the directionof each feature point is calculated Finally the SIFT featurevector of each key point is generated In the process ofgenerating SIFT feature vector 16 seed points are selectedfor each key point and each seed point has 8 orientationsTherefore the 16 times 8 = 128 dimensions of the SIFT featurevector can be obtained for each key point After obtainingthe SIFT feature vectors of the Gaussian curvature featurematrices corresponding to the two adjacent depth images thetwoGaussian curvature featurematrices can bematched andthe similarity of the vectors ismeasured byEuclidean distancein the process of matching
After the key points of the Gaussian curvature featurematrices corresponding to the two adjacent depth images arematched the motion equation of visual object can be esti-mated by using the pixelsrsquo 3D coordinates which are obtainedby antiprojection transformation The motion equation isdefined as
x1015840119897= Rx119897+ T (6)
where x119897and x1015840119897are respectively the 3D coordinates of points
before and after visual objectmotionR is the unit orthogonalrotation matrix and T is the translation vector then theresult of motion estimation is the solution of minimizing theobjective function 119891(RT) The objective function 119891(RT) isdefined as
119891 (RT) =119872
sum
119897=1
10038171003817100381710038171003817x1015840119897minus (Rx
119897+ T)10038171003817100381710038171003817
2
(7)
where119872 is the number of key points which are matched Inorder to solve the R and T in formula (7) the matrix H isdefined as
H =
119872
sum
119897=1
x1015840clx119879
cl (8)
where x1015840cl = x1015840119897minus (1119872)sum
119872
119897=1x1015840119897 xcl = x
119897minus (1119872)sum
119872
119897=1x119897
By carrying out singular value decomposition on H H canbe expressed as H=UΛV119879 then according to U and V R informula (7) can be deduced as
R = VU119879 (9)
After obtaining R substitute R into formula (10) tocalculate T namely
T = x1015840119897minus Rx119897 (10)
After obtaining R and T the motion equation shown informula (6) can be obtained and themotion estimation of thevisual object is accomplished
33 Planning the Next View Based on Motion Estimationand Best View Model After the motion equation of visualobject is estimated the next view of the camera can beplanned Analysis can be known if substituting the cameraobservation position into the motion equation of visualobject the camera can do the synchronous motion with thevisual object In addition on the basis of the synchronousmotion with the visual object if the camera motion offsetfor avoiding self-occlusion is introduced the whole processof dynamic self-occlusion avoidance can be achieved Byanalyzing the best view model in formula (5) we can knowit is the camera best view that can make formula (11) achievemaximum value Formula (11) is defined as
119891 (x) =119873
sum
119895=1
119878119895cos (120579
119895 (x)) (11)
where x is the camera observation position Since formula (11)is a nonconvex model it is difficult to directly calculate itsglobal optimal solution Meanwhile because the problem ofdynamic self-occlusion avoidance needs to calculate a seriesof next views and the gradient information in formula (11) canprovide basis for determining the cameramotion offset of theself-occlusion avoidance process in the process of dynamicself-occlusion avoidance the gradient descent idea is usedin this paper for defining the formula of camera observationposition x1015840
119873119881as
x1015840119873119881
= Rx119881+ T + 120575nabla119891 (x
119881) (12)
where x119881is the current observation position of the camera
and 120575 is the offset coefficient Based on the solution to formula(12) the formula of next observation direction k
119873119881is defined
as
k119873119881
= xmid minus x1015840119873119881
(13)
According to the constraint condition in formula (5) wecan search one point x
119873119881along the opposite direction of k
119873119881
and make x119873119881
meet1003817100381710038171003817xmid minus x
119873119881
10038171003817100381710038172 =10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172 (14)
By now the obtained view (x119873119881
v119873119881
) can be regardedas the new current view (x
119881 v119881) After acquiring the depth
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Mathematical Problems in Engineering 5
method in literature [12] is used for calculating the Gaussiancurvature feature matrix of depth image
322 Estimating the Motion Equation of Visual Object Basedon the Matched Feature Points Analysis shows that thematched points can provide reliable information for themotion estimation if the Gaussian curvature feature matricescorresponding to two adjacent depth images of the visualobject can be matched SIFT algorithm in literature [13]is more efficient and it is not only invariant to imagetranslation rotation and scaling but also robust to thechange of visual angle affine transformation and noise sothe SIFT algorithm is adopted in this paper for matching thefeature points
SIFT algorithm mainly consists of two steps the gener-ation of SIFT features and the matching of feature vectorsIn order to generate the SIFT features firstly the scale spaceshould be established Because the Difference of Gaussiansscale-space (DoG scale-space) has good stability in detectionof key points the DoG scale-space is used for detecting thekey points in this paper Secondly the key points which areinstable and sensitive to noise are filtered Then the directionof each feature point is calculated Finally the SIFT featurevector of each key point is generated In the process ofgenerating SIFT feature vector 16 seed points are selectedfor each key point and each seed point has 8 orientationsTherefore the 16 times 8 = 128 dimensions of the SIFT featurevector can be obtained for each key point After obtainingthe SIFT feature vectors of the Gaussian curvature featurematrices corresponding to the two adjacent depth images thetwoGaussian curvature featurematrices can bematched andthe similarity of the vectors ismeasured byEuclidean distancein the process of matching
After the key points of the Gaussian curvature featurematrices corresponding to the two adjacent depth images arematched the motion equation of visual object can be esti-mated by using the pixelsrsquo 3D coordinates which are obtainedby antiprojection transformation The motion equation isdefined as
x1015840119897= Rx119897+ T (6)
where x119897and x1015840119897are respectively the 3D coordinates of points
before and after visual objectmotionR is the unit orthogonalrotation matrix and T is the translation vector then theresult of motion estimation is the solution of minimizing theobjective function 119891(RT) The objective function 119891(RT) isdefined as
119891 (RT) =119872
sum
119897=1
10038171003817100381710038171003817x1015840119897minus (Rx
119897+ T)10038171003817100381710038171003817
2
(7)
where119872 is the number of key points which are matched Inorder to solve the R and T in formula (7) the matrix H isdefined as
H =
119872
sum
119897=1
x1015840clx119879
cl (8)
where x1015840cl = x1015840119897minus (1119872)sum
119872
119897=1x1015840119897 xcl = x
119897minus (1119872)sum
119872
119897=1x119897
By carrying out singular value decomposition on H H canbe expressed as H=UΛV119879 then according to U and V R informula (7) can be deduced as
R = VU119879 (9)
After obtaining R substitute R into formula (10) tocalculate T namely
T = x1015840119897minus Rx119897 (10)
After obtaining R and T the motion equation shown informula (6) can be obtained and themotion estimation of thevisual object is accomplished
33 Planning the Next View Based on Motion Estimationand Best View Model After the motion equation of visualobject is estimated the next view of the camera can beplanned Analysis can be known if substituting the cameraobservation position into the motion equation of visualobject the camera can do the synchronous motion with thevisual object In addition on the basis of the synchronousmotion with the visual object if the camera motion offsetfor avoiding self-occlusion is introduced the whole processof dynamic self-occlusion avoidance can be achieved Byanalyzing the best view model in formula (5) we can knowit is the camera best view that can make formula (11) achievemaximum value Formula (11) is defined as
119891 (x) =119873
sum
119895=1
119878119895cos (120579
119895 (x)) (11)
where x is the camera observation position Since formula (11)is a nonconvex model it is difficult to directly calculate itsglobal optimal solution Meanwhile because the problem ofdynamic self-occlusion avoidance needs to calculate a seriesof next views and the gradient information in formula (11) canprovide basis for determining the cameramotion offset of theself-occlusion avoidance process in the process of dynamicself-occlusion avoidance the gradient descent idea is usedin this paper for defining the formula of camera observationposition x1015840
119873119881as
x1015840119873119881
= Rx119881+ T + 120575nabla119891 (x
119881) (12)
where x119881is the current observation position of the camera
and 120575 is the offset coefficient Based on the solution to formula(12) the formula of next observation direction k
119873119881is defined
as
k119873119881
= xmid minus x1015840119873119881
(13)
According to the constraint condition in formula (5) wecan search one point x
119873119881along the opposite direction of k
119873119881
and make x119873119881
meet1003817100381710038171003817xmid minus x
119873119881
10038171003817100381710038172 =10038171003817100381710038171003817xmid minus xbegin
100381710038171003817100381710038172 (14)
By now the obtained view (x119873119881
v119873119881
) can be regardedas the new current view (x
119881 v119881) After acquiring the depth
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
6 Mathematical Problems in Engineering
image of visual object in the new view (x119881 v119881) the process
from formula (6) to formula (14) is repeated until thedifference between two 119891(x) which are calculated in twoadjacent views according to formula (11) is less than a giventhreshold It means that a series of views (x
119873119881 v119873119881
) can becalculated during the visual object motion so the process ofdynamic self-occlusion avoidance can be accomplished
34 The Algorithm of Dynamic Self-Occlusion Avoidance
Algorithm 1 (dynamic self-occlusion avoidance)
Input The camera internal and external parameters and twoadjacent depth images acquired in initial camera view
Output The set of a series of camera views corresponding tothe dynamic self-occlusion avoidance
Step 1 Calculate the pixelsrsquo 3D coordinates and the Gaussiancurvature feature matrices corresponding to the two adjacentdepth images
Step 2 Detect the self-occlusion boundary in the seconddepth image and establish the self-occlusion region in thesecond depth image according to formula (1) to formula (4)
Step 3 Construct the best view model according to self-occlusion information (formula (5))
Step 4 Match the key points based on Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages by using SIFT algorithm
Step 5 Based on the 3D coordinates of matched key pointssolve the objective function in formula (7) according toformula (8) to formula (10) then the visual objectrsquos motionequation in formula (6) can be obtained
Step 6 Plan the next view of camera according to formula (11)to formula (14)
Step 7 In the obtained next view acquire a depth image andcalculate 119891(x) in formula (11)
Step 8 If the difference between two adjacent 119891(x) is lessthan a given threshold the process of dynamic self-occlusionavoidance can be terminated and otherwise jump to Step 4
to continue the process of dynamic self-occlusion avoidance
4 Experiments and Analysis
41 Experimental Environment In order to validate the effectof the method in this paper the OpenGL is adopted toimplement the process of dynamic self-occlusion avoidanceduring camera observing the moving object based on 3Dobject model in the Stuttgart Range Image Database Theexperimental hardware environment is Intel CPU (R) Core(TM) i7-3770 340GHz and the memory is 800GB Thedynamic self-occlusion avoidance program is implemented
with C++ language In the process of self-occlusion avoid-ance the parameter of the projection matrix in OpenGL is(60 1 200 600) and the window size is 400 times 400
42 The Experimental Results and Analysis To verify thefeasibility of the proposed method we do experiments withdifferent visual objects in different motion modes Duringthe experiments the initial camera observation position isxbegin = [000 minus100 30000]
119879 and the initial observationdirection is kbegin = [000 100 minus30000]
119879 The coordinateunit is millimeter (mm) and the offset coefficient is 120575 = 4Theprocess of dynamic self-occlusion avoidance will terminate ifthe calculated difference of 119891(x) between two adjacent viewsis less than the given threshold 30mm2 The offset coefficientand threshold are chosen by empirical value Due to thelimited space here we show 3 groups of experimental resultscorresponding to the visual object Bunny with relatively largeself-occlusion region and Duck with relatively small self-occlusion region The results are shown in Tables 1 2 and3 respectively Table 1 shows the process of dynamic self-occlusion avoidance in which the visual object Bunny doestranslation along [1 0 0]119879 with the speed of 3mms Table 2shows the process of dynamic self-occlusion avoidance inwhich the visual object Bunnydoes translation along [2 1 0]119879
with the speed ofradic5mms and rotation around [minus4 1 25]119879with the speed of 1∘s Table 3 shows the process of dynamicself-occlusion avoidance inwhich the visual objectDuck doestranslation along [2 1 0]
119879 with the speed of radic5mms androtation around [minus4 1 25]
119879 with the speed of 1∘s Becausethe camera needs multiple motions in the process of dynamicself-occlusion avoidance only partial observing results thematched results of Gaussian curvature feature matrices andrelated camera information are shown in Tables 1 2 and 3after the results of first three images are shown
The first row in Tables 1 2 and 3 shows the depth imagenumber corresponding to different views in the process ofself-occlusion avoidance At the end of the self-occlusionavoidance the image number in Tables 1 2 and 3 is 1716 and 13 respectively Thus we can see that the cameramovement times are different for different visual objects orthe same visual object with different motions in the processof self-occlusion avoidance The second row in Tables 1 2and 3 shows the acquired depth images of visual object indifferent camera views which are determined by proposedmethod From the images in this row we can see that theself-occlusion region is gradually observed more and morein the process of self-occlusion avoidance for example theself-occlusion region mainly caused by Bunnyrsquos ear in thesecond image in Tables 1 and 2 and the self-occlusion regionmainly caused by Duckrsquos mouth in the second image inTable 3 are better observed when the process of self-occlusionavoidance is accomplished The third row in Tables 1 2and 3 shows the matched results of the Gaussian curvaturefeature matrices corresponding to the two adjacent depthimages with SIFT algorithm These matched results are thebasis of motion estimation and self-occlusion avoidance andgood matched results are helpful to obtain accurate motionestimation then the accurate camera motion trajectory can
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Mathematical Problems in Engineering 7
Table 1 The self-occlusion avoidance process of visual object Bunny with translation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 10 sdot sdot sdot 16 17
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 221 149 sdot sdot sdot 214 sdot sdot sdot 305 sdot sdot sdot 323 mdash
Error matchedpoints mdash 32 33 sdot sdot sdot 47 sdot sdot sdot 55 sdot sdot sdot 35 mdash
Error rate mdash 145 222 sdot sdot sdot 219 sdot sdot sdot 180 sdot sdot sdot 108 mdash
Observationposition
000 minus10030000
000 minus10030000
6005 minus37129306 sdot sdot sdot
15494minus602024336
sdot sdot sdot
22252minus1040316734
sdot sdot sdot25637
minus12216 941326057
minus12269 8872
Observationdirection
000 100minus30000
000 100minus30000
minus5456 minus173minus29499 sdot sdot sdot
minus16823 9591minus22913 sdot sdot sdot
minus2459712077minus12212
sdot sdot sdotminus25867
12317 minus8900minus26593
12399 minus6251
Observed area mdash 16761 58702 sdot sdot sdot 215860 sdot sdot sdot 262568 sdot sdot sdot 301434 302558
be further calculated Therefore the matched results in thethird row will influence the effect of self-occlusion avoidanceto a certain extent The fourth fifth and sixth rows in Tables1 2 and 3 respectively show the total matched points theerror matched points and the error rate in the process ofmatching with SIFT algorithm where the error rate is equalto the error matched points divided by the total matchedpointsThe seventh eighth and ninth rows in Tables 1 2 and3 respectively show the observation position observationdirection and the observed area of the vested self-occlusionregion in related view when the visual object Bunny andDuck do different motions It can be seen from these dataand their change trends that the process of avoiding self-occlusion dynamically with proposed method accords withthe observing habit of human vision
In view of the fact that there is nomethod of dynamic self-occlusion avoidance to be compared and no Ground Truthabout dynamic self-occlusion avoidance currently it is verydifficult to evaluate our method by comparing with othermethods or Ground Truth In order to make the furtherquantitative analysis of the feasibility and effectiveness of ourmethod after deeply analyzing the characteristics of dynamicself-occlusion avoidance we propose an index named ldquoeffec-tive avoidance raterdquo to measure the performance of dynamicself-occlusion avoidance method Considering the effect ofobject surface the same object will have different self-occlusion regions in different camera views and meanwhilethe distribution of the self-occlusion region is random it is
not realistic to find a camera view in which the camera canobtain all the information of the self-occlusion region so itis inappropriate to take the ratio of the final observed self-occlusion region area to the total self-occlusion region area asevaluation criterion In addition because the goal of dynamicself-occlusion avoidance is that the camera not only can trackthe visual object but also can observe the vested self-occlusionregion maximally the camera will inevitably gradually beclose to the best view in the process of dynamic self-occlusionavoidance namely the objective view described in formula(5) The area in the objective view xBV is also the self-occlusion region area really expected to be observed Basedon this the proposed index of effective avoidance rate isdefined as
120578 =
119878viewsum119873
119895=1119878119895
119878purposesum119873
119895=1119878119895
times 100 =119878view119878purpose
times 100 (15)
where 120578 is the effective avoidance rate 119878view is the observedself-occlusion region area in the view where the process ofdynamic self-occlusion avoidance is accomplished 119878purposeis the self-occlusion region area in the objective viewand sum
119873
119895=1119878119895is the total self-occlusion region area to be
avoidedIn order to make quantitative analysis of the performance
of proposed method based on the effective avoidance ratewe have done multiple groups of experiments by adjust-ing the motion modes of different visual objects Table 4
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
8 Mathematical Problems in Engineering
Table 2 The self-occlusion avoidance process of visual object Bunny with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 5 sdot sdot sdot 9 sdot sdot sdot 15 16
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 224 162 sdot sdot sdot 205 sdot sdot sdot 247 sdot sdot sdot 338 mdash
Error matchedpoints mdash 20 11 sdot sdot sdot 30 sdot sdot sdot 42 sdot sdot sdot 29 mdash
Error rate mdash 89 68 sdot sdot sdot 146 sdot sdot sdot 170 sdot sdot sdot 86 mdash
Observationposition
000 minus10030000
000 minus10030000
5668 minus20029507 sdot sdot sdot
12716 minus457426626 sdot sdot sdot
21141 minus977818011 sdot sdot sdot
25384minus1114710268
26232minus11725 9219
Observationdirection
000 100minus30000
000 100minus30000
minus5909 minus485minus29409 sdot sdot sdot
minus133205028minus26407
sdot sdot sdot
minus2264012939minus14833
sdot sdot sdotminus25515
13429 minus8287minus25698
14529 minus5343
Observed area mdash 16761 62548 sdot sdot sdot 145364 sdot sdot sdot 243137 sdot sdot sdot 264857 265463
Table 3 The self-occlusion avoidance process of visual object Duck with both translation and rotation
Information Image number1 2 3 sdot sdot sdot 6 sdot sdot sdot 9 sdot sdot sdot 12 13
Depth image sdot sdot sdot sdot sdot sdot sdot sdot sdot
MatchingGaussiancurvaturefeature matriceswith SIFTalgorithm
mdash sdot sdot sdot sdot sdot sdot sdot sdot sdot mdash
Total matchedpoints mdash 552 431 sdot sdot sdot 456 sdot sdot sdot 432 sdot sdot sdot 438 mdash
Error matchedpoints mdash 20 20 sdot sdot sdot 38 sdot sdot sdot 22 sdot sdot sdot 71 mdash
Error rate mdash 36 46 sdot sdot sdot 83 sdot sdot sdot 51 sdot sdot sdot 162 mdashObservationposition
000 minus10030000
000 minus10030000
minus018 261329759 sdot sdot sdot
minus932 656229207 sdot sdot sdot
minus198810603 27918 sdot sdot sdot
minus497413568 26574
minus243010582 25845
Observationdirection
000 100minus30000
000 100minus30000
084 minus2599minus29887 sdot sdot sdot
1472 minus6807minus29181 sdot sdot sdot
1758 minus9239minus28488 sdot sdot sdot
2396minus12071minus27360
5605minus11493minus27139
Observed area mdash 8304 22232 sdot sdot sdot 31765 sdot sdot sdot 46284 sdot sdot sdot 70577 73358
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Mathematical Problems in Engineering 9
Table 4 The quantitative analysis results for 10 groups of experiments
Groupnumber
Visualobject Initial camera view Motion mode
sum119873
119895 = 1119878119895
(mm2)119878purpose(mm2)
119878view(mm2)
120578
()119879
(s) 119873119879
(s)
1 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879Translation along[1 0 0]119879 with thespeed of 3mms
728926 318912 302558 9487 158 17 093
2 Bunnyxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
728926 318912 265463 8324 155 16 097
3 Duckxbegin = [000 minus100 30000]119879
vbegin = [000 100 minus30000]119879
Translation along[2 1 0]119879 with thespeed ofradic5mms
and rotationaround [minus4 125]119879 with thespeed of 1∘s
191110 102527 73358 7155 133 13 102
4 Bunnyxbegin = [12000 10000 minus25600]119879
vbegin = [minus12000 minus10000 minus25600]119879Rotation around[7 minus3 4]119879 with the
speed of 2∘s502892 323847 301748 9318 136 15 091
5 Bunnyxbegin = [15000 17000 19600]119879
vbegin = [minus15000 minus17000 minus19600]119879
Translation along[1 2 1]119879 with thespeed ofradic6mms
and rotationaround [2 minus1 minus2]119879with the speed of
1∘s
441450 175818 155348 8836 171 18 095
6 Duckxbegin = [10000 minus15000 24000]119879
vbegin = [minus10000 15000 minus24000]119879
Translation along[3 2 minus1]119879 with the
speed ofradic14mms
330746 278837 274835 9856 195 21 093
7 Molexbegin = [minus21000 5000 21000]119879
vbegin = [21000 minus5000 minus21000]119879
Translation along[2 0 2]119879 with thespeed ofradic8mms
and rotationaround [1 minus1 minus2]119879with the speed of
3∘s
32001 22168 13309 6004 112 11 102
8 Rockerxbegin = [minus19600 17000 15000]119879
vbegin = [19600 minus17000 minus15000]119879Translation along[1 1 2]119879 with thespeed ofradic6mms
190509 144092 136345 9462 84 9 093
9 Rockerxbegin = [19000 minus20000 11800]119879
vbegin = [minus19000 20000 minus11800]119879
Translation along[2 minus1 minus1]119879 withthe speed ofradic6mms androtation around[minus1 2 minus3]119879 withthe speed of 2∘s
343354 270733 194798 7195 98 10 098
10 Dragonxbegin = [19000 20000 11800]119879
vbegin = [minus19000 minus20000 minus11800]119879
Translation along[2 4 minus1]119879 with the
speed ofradic21mms androtation around[minus1 2 minus2]119879 withthe speed of 3∘s
573648 194861 148598 7626 165 15 110
Average mdash mdash mdash 406356 215071 186636 8678 141 145 097
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
10 Mathematical Problems in Engineering
shows the quantitative analysis results of total 10 groupsof experiments which include the 3 groups of experimentsdescribed above and the other 7 groups In addition in thelast column of each experiment the average time whichis from the time that the camera arrives at a certain viewto the time that the next view has been calculated isshown
From the 10 groups of experiments in Table 4 we can seethat affected by surface of visual object the different self-occlusion region area sum119873
119895=1119878119895will be generated for the same
(or different) visual object in different (or the same) cameraviews and the distribution of the self-occlusion region israndom so the vested self-occlusion region area sum119873
119895=1119878119895will
not be observed completely even if the camera is in thebest view (the objective view) At the same time the cameracan only infinitely be close to but hardly reach the objectiveview accurately in the process of self-occlusion avoidancethat is to say at the end of self-occlusion avoidance thecamera observed area 119878view can only be close to but notreach the area 119878purpose in the objective view which alsoverifies the rationality of taking the effective avoidance rate120578 as evaluation criterion Besides comparing 10 groups ofexperiments in Table 4 we can know that the visual objectsin groups 1 4 6 and 8 only do translation or rotation indifferent initial views and the motion is relatively simplethen the average effective avoidance rate of these fourgroups is 9529 But in the groups 2 3 5 7 9 and 10 inTable 4 the visual objects do both translation and rotationin different initial views the motion is relatively complexand then the average effective avoidance rate of these sixgroups is 7842 which is obviously lower than that of thefour preceding groups in which the visual objects only dotranslation or rotation The main reason from our analysisis that many steps in the process of dynamic self-occlusionavoidance such as antiprojection transformation matchingpoints based on SIFT andmotion estimation will make errorand the error will accumulate when the camera changes itsview The most important thing is that error accumulationwill become more obvious when the motion of the visualobject is more complex which will lead to the result thatthe effective avoidance rate will be significantly lower whenthe motion of the visual object is complex Objectivelyspeaking it is inevitable that there are error and erroraccumulation in the process of self-occlusion avoidanceand it is not realistic to eliminate the effect of these errorscompletely what different methods can do is to reduce orcut down the effect of these errors The last row data inTable 4 shows that the average effective avoidance rate ofthe proposed method in this paper reaches 8678 thereforein the average sense the better result of dynamic self-occlusion avoidance can be obtained based on the proposedmethod Furthermore from the last column data in Table 4it can be seen that the average time consumption is 097swhich is from the time that the camera arrives at a certainview to the time that the next view has been calculatedIt shows that the proposed method has good real-timeperformance
5 Conclusion
In this paper an approach for avoiding self-occlusion dynam-ically is proposed based on the depth image sequence ofmoving visual object The proposed approach does not needspecific equipment and the a priori knowledge of visual objector limit the observation position of the camera on a fixedsurface This work is distinguished by three contributionsFirstly the problem of dynamic self-occlusion avoidancewhich is still in initial stage at present is studied and thesolution to dynamic self-occlusion avoidance problem is pro-posed and implemented which provides an idea for furtherstudying the dynamic self-occlusion avoidance Secondlybased on the depth images of moving visual object theGaussian curvature feature matrices are matched by SIFTalgorithm and the motion equation is efficiently estimatedby using singular value decompositionThis provides a novelidea about motion estimation in 3D space based on the depthimage of visual object Thirdly considering that there are noevaluation criterions about dynamic self-occlusion avoidanceat present an evaluation criterion named ldquoeffective avoidanceraterdquo is proposed to measure the performance of dynamicself-occlusion avoidance reasonably and it provides thereferable evaluation basis for further studying the problem ofdynamic self-occlusion avoidance The experimental resultsshow that the proposed approach achieves the goal thatthe camera avoids the self-occlusion automatically when thevisual object is moving and the avoidance behavior accordswith the observing habit of human vision
Competing Interests
The authors declare that there is no conflict of interestsregarding the publication of this article
Acknowledgments
This work is supported by the National Natural ScienceFoundation of China under Grant no 61379065 and theNatural Science Foundation of Hebei province in Chinaunder Grant no F2014203119
References
[1] C Connolly ldquoThe determination of next best viewsrdquo in Pro-ceedings of the 1985 IEEE International Conference on Roboticsand Automation pp 432ndash435 IEEEMissouriMo USAMarch1985
[2] R Pito ldquoA sensor-based solution to the lsquonext best viewrsquoproblemrdquo in Proceedings of the 13th International Conferenceon Pattern Recognition (ICPR rsquo96) vol 1 pp 941ndash945 IEEEVienna Austria August 1996
[3] P Whaite and F P Ferrie ldquoAutonomous exploration drivenby uncertaintyrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 19 no 3 pp 193ndash205 1997
[4] Y F Li and Z G Liu ldquoInformation entropy-based viewpointplanning for 3-D object reconstructionrdquo IEEE Transactions onRobotics vol 21 no 3 pp 324ndash337 2005
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Mathematical Problems in Engineering 11
[5] M Trummer C Munkelt and J Denzler ldquoOnline next-best-view planning for accuracy optimization using an extended E-criterionrdquo in Proceedings of the 20th International Conference onPattern Recognition (ICPR rsquo10) pp 1642ndash1645 IEEE IstanbulTurkey August 2010
[6] J E Banta L M Wong C Dumont and M A Abidi ldquoA next-best-view system for autonomous 3-D object reconstructionrdquoIEEE Transactions on Systems Man and Cybernetics Part ASystems and Humans vol 30 no 5 pp 589ndash598 2000
[7] W Fang and B He ldquoAutomatic view planning for 3D recon-struction and occlusion handling based on the integrationof active and passive visionrdquo in Proceedings of the 21st IEEEInternational Symposium on Industrial Electronics (ISIE rsquo12) pp1116ndash1121 Hangzhou China May 2012
[8] J I Vasquez-Gomez L E Sucar and R Murrieta-Cid ldquoHierar-chical ray tracing for fast volumetric next-best-view planningrdquoin Proceedings of the 10th International Conference on Computerand Robot Vision (CRV rsquo13) pp 181ndash187 IEEE SaskatchewanCanada May 2013
[9] C Potthast and G S Sukhatme ldquoA probabilistic framework fornext best view estimation in a cluttered environmentrdquo Journalof Visual Communication and Image Representation vol 25 no1 pp 148ndash164 2014
[10] B Wu X Sun Q Wu M Yan H Wang and K Fu ldquoBuildingreconstruction from high-resolution multiview aerial imageryrdquoIEEE Geoscience and Remote Sensing Letters vol 12 no 4 pp855ndash859 2015
[11] S Zhang and J Liu ldquoA self-occlusion detection approach basedon depth image using SVMrdquo International Journal of AdvancedRobotic Systems vol 9 article 230 2012
[12] P J Besl and R C Jain ldquoSegmentation through variable-order surface fittingrdquo IEEE Transactions on Pattern Analysis andMachine Intelligence vol 10 no 2 pp 167ndash192 1988
[13] N Bayramoglu and A A Alatan ldquoShape index SIFT rangeimage recognition using local featuresrdquo in Proceedings of the20th International Conference on Pattern Recognition (ICPR rsquo10)pp 352ndash355 IEEE Istanbul Turkey August 2010
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of
Submit your manuscripts athttpwwwhindawicom
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical Problems in Engineering
Hindawi Publishing Corporationhttpwwwhindawicom
Differential EquationsInternational Journal of
Volume 2014
Applied MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Probability and StatisticsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Mathematical PhysicsAdvances in
Complex AnalysisJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
OptimizationJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
CombinatoricsHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Operations ResearchAdvances in
Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Function Spaces
Abstract and Applied AnalysisHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
International Journal of Mathematics and Mathematical Sciences
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Algebra
Discrete Dynamics in Nature and Society
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Decision SciencesAdvances in
Discrete MathematicsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014 Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Stochastic AnalysisInternational Journal of