icbv course final project arik krol aviad pinkovezky

16
ICBV Course Final Project Arik Krol Aviad Pinkovezky

Upload: paula-james

Post on 23-Dec-2015

218 views

Category:

Documents


4 download

TRANSCRIPT

ICBV Course Final ProjectArik Krol

Aviad Pinkovezky

Motivation:

• Current MMI is based mainly on “Point & Click”

devices

• Video Capturing as a new Approach for MMI

• Hardware is available - Web Cameras and

powerful

processors

• Potential Usages – Working with laptops, users

with

“hands on keyboards”, etc

Goals:

• Exploring the field of motion detection

• Exploring the field of skin colors distinction

• A working Demo that can detect palm

movements:

• No Real Time, yet…

• Minimal rate of False detections

• Determine Direction of movement

Motion Detection:

• First approach: Segmentation by Clustering (K-

Means)

• Motion Detection by tracking the centers of

gravity of

clusters over the frames

• The Problem – Complexity of Calculation, doesn’t

fit into

real time scenario!

Motion Detection (cont.):

• Second approach: Subtracting consecutive

frames

• Motion Detection by tracking the difference in

pixels

values

• Note - Assumptions are: Relatively static

background and

stationary camera

Motion Detection implementation:• For each two consecutive frames:

• Convert from RGB to Grayscale

• Subtraction

• Gaussian Smoothing

Skin Color Detection:

• H.S.V – Hue, Saturation, Value

• An alternative representation of color pixels

• Enables us to isolate Hue levels, regardless of

Saturation

and Value levels

Skin Color Detection (cont.):

• The human skin is characterized by different

levels of red

hue - 335 to 25 degrees

• Value level is greater than 40

Skin Motion Detection:

• { Motion Pixels } { Skin Pixels} = {Skin

Motion Pixels}

• Direction of movement – Determined by the

differences of

X axis value averages between consecutive

frames

• Setting adequate thresholds – by trial and error

Skin Motion Detection:

• Now, let’s try to detect a moving piece of paper:

• Skin Motion detection results finally in no

detection at all.

Problems we encountered:

• Face – Can create false detection of skin movement (head

movements & non

skin movement) – Solved by tracking the 1/3 bottom part of the

image.

• Complexity of calculation – better than clustering, yet not real

time

like – unsolved

• Skin like objects – may cause false detection

A Few Results:

Future Improvements:

• Improving run time performances to support real

time

motion detection, can be achieved by:

Using different programming languages

Using hardware acceleration (parallel

computing,

GPGPU, etc.)

• Setting thresholds dynamically by calibrating

the system.

• Identifying a larger variety of movements, and

adding new

features accordingly

References:

• Francesca Gasparini, Raimondo Schettini, Skin

segmentation using

multiple thresholdings

• University of Sussex, UK. Web page of David Young,

“Static Camera

and moving objects”:

http://www.cogs.susx.ac.uk/users/davidy/teachvision/vi

sion6.html#heading3

• And of course, Wikipedia – H.S.V