dinesh manocha university of maryland at college park dm ...dinesh manocha university of maryland at...
TRANSCRIPT
![Page 1: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/1.jpg)
Dinesh ManochaUniversity of Maryland at College Park
Prediction of Human Motion & Traffic Agents in Dense Environments
![Page 2: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/2.jpg)
Collaborators
2
• Aniket Bera (UNC/UMD)• Uttaran Bhattacharya (UMD)• Rohit Chandra (UMD)• Ernest Cheung (UMD)• Kurt Gray (UNC)• Sujeong Kim (UNC/SRI)• Yuexin Ma (UNC/Baidu/HKU)• Tanmay Randhavane (UNC)
![Page 3: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/3.jpg)
Real-World Pedestrian/Crowd Analysis
• Behavior Learning•Culture characteristics•Crowd prediction
![Page 4: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/4.jpg)
Analyze Crowd Movements
![Page 5: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/5.jpg)
Driverless Cars: Pedestrian Interaction
Source: Oxbotica at Oxford University
![Page 6: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/6.jpg)
Current AD technology vs. Real-world Scenarios
• Many traffic situations are still too challenging for autonomous vehicles
6
Current Autonomous Driving Urban Traffic Condition: China
![Page 7: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/7.jpg)
Challenging Traffic Conditions: China
7Current technologies and datasets for dense traffic are limited
![Page 8: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/8.jpg)
More Challenging Conditions: India
8
No respect for rules; cultural norms, driver & pedestrian behaviors
![Page 9: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/9.jpg)
Organization
9
• Pedestrian and Crowd Motions
• Heterogeneous multi-agent simulation
• Tracking urban traffic & Prediction
• Driver behavior modeling
![Page 10: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/10.jpg)
Organization
10
• Pedestrian and Crowd Motions
• Heterogeneous multi-agent simulation
• Tracking urban traffic & Prediction
• Driver behavior modeling
![Page 11: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/11.jpg)
Pedestrian and Crowd Motion: Tracking & Prediction11
• New motion models based on RVO (reciprocal velocity obstacles)
• Combine motion model with behavior models
• Real time tracking: deep learning + motion models
• Learning Pedestrian Dynamics using Bayesian Inferences
• Handling Dense Crowds
![Page 12: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/12.jpg)
Realtime Pedestrian Tracking in Dense Crowds
[Chandra et al. 2019]
![Page 13: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/13.jpg)
REALTIME TRACKING AND PERSONALITY MODELING
[Bera et al. 2017]
![Page 14: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/14.jpg)
Pedestrian/Crowd Movement Prediction
[Bera et al. 2016, 2017]
![Page 15: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/15.jpg)
CROWD BEHAVIOR MOVEMENT PREDICTION
![Page 16: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/16.jpg)
2017 PRESIDENTIAL INAUGURATION CROWDS
[Bera et al. 2017]
![Page 17: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/17.jpg)
Data Driven Crowd Simulation & Prediction
[Kim et al. 2017
![Page 18: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/18.jpg)
STEP: Spatial Temporal Graph Convolutional Networks for Emotion Perception from Gaits
![Page 19: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/19.jpg)
Can you Classify Emotions from Movements?
19
!"#$%&$ = −0.67-./01"# = −0.25
!"#$%&$ = 0.53-./01"# = −0.11
!"#$%&$ = 0.10-./01"# = −0.07
-./01"# = 0.68!"#$%&$ = −0.05
Which is angry, sad, happy, neutral?
![Page 20: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/20.jpg)
Emotion Classification from Pedestrian Motion
20
!"#$%&$ = −0.67-./01"# = −0.25
!"#$%&$ = 0.53-./01"# = −0.11
!"#$%&$ = 0.10-./01"# = −0.07
-./01"# = 0.68!"#$%&$ = −0.05
Which is angry, sad, happy, neutral?Forthcoming E-Walk Dataset
[Randhavane et al. 2019]
![Page 21: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/21.jpg)
Organization
22
• Pedestrian and Crowd Motions
• Heterogeneous multi-agent simulation
• Tracking urban traffic & Prediction
• Driver behavior modeling
![Page 22: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/22.jpg)
Dense traffic scenarios: Heterogeneous Agents
Vehicles (big and small), pedestrians, bicycles, tricycles, etc
![Page 23: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/23.jpg)
Heterogeneous Multi-Agent Navigation
��
Agents:
• Varying shapes
• Varying dynamics
• Different behaviors
• Operating in tight spaces
![Page 24: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/24.jpg)
�$��� �����!��������#��� �"������&��&%������'
��������
��������
Heterogeneous Multi-Agent Representation
![Page 25: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/25.jpg)
Kinematic Models: Different Agents
��������
Car tricycle bicycle pedestrian
Simple Car Model [Laumond et al. 1998]
![Page 26: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/26.jpg)
AutoRVO: Preferred Steering Computation
��������
�$��� �����!��������#��� �"������&��&%������'
��������
��������
Search for free-space for collision-free local navigation
![Page 27: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/27.jpg)
AutoRVO: Results
��������
![Page 28: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/28.jpg)
Comparisons: Multi-Agent Navigation
![Page 29: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/29.jpg)
Organization
30
• Pedestrian and Crowd Motions
• Heterogeneous multi-agent simulation
• Tracking urban traffic & Prediction
• Driver behavior modeling
![Page 30: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/30.jpg)
Highlights
• High degree of heterogeneity
• Dense Traffic
• No traffic protocols in place
![Page 31: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/31.jpg)
Consists of 4 stages
ApproachCombine model-based and learning-based methods
![Page 32: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/32.jpg)
Stage 1 – Agent Detection Using Mask R-CNN
Use Mask R-CNN based agent detection to generate Segmented Boxes of each agent
![Page 33: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/33.jpg)
Stage 2 – Velocity Prediction Using HTMI
Use HTMI to model inter-agent interactions and collision avoidance
HTMI: Heterogeneous motion model
![Page 34: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/34.jpg)
Stage 3 – Feature Extraction Using Segmented Boxes
Generate novel features called “Deep TA-features” from segmented boxes
![Page 35: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/35.jpg)
Stage 4 – Feature Matching Using IOU Overlap
We use the cosine distance metric
![Page 36: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/36.jpg)
Results – Low Density Traffic
Agents: Cars, Scooters, Bikes, Rickshaws, Pedestrians
Car: Green
Pedestrians & Two-Wheelers: Red
Rickshaws: Purple
![Page 37: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/37.jpg)
Results – Medium Density
Agents: Cars, Two-Wheelers, Rickshaws, Pedestrians
Car: Green
Pedestrians & Two-Wheelers: Red
Buses: Cyan
Agents: Cars, Bicycles, Pedestrians, Buses
![Page 38: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/38.jpg)
Results – High Density
Agents: Cars, Two-Wheelers, Rickshaws, PedestriansAgents: Cars, Scooters, Bicycles, Rickshaws, Pedestrians, Animals
Car: Green
Rickshaws: Purple
Pedestrians & Two-Wheelers: Red
Buses: Cyan
Animals: Yellow
![Page 39: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/39.jpg)
Strengths – IWe can track drivers inside different road agents
![Page 40: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/40.jpg)
Strengths – IIWe can track atypical agents
![Page 41: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/41.jpg)
Strengths – IIIWe can track agents in challenging conditions
• Night time with a jittery, moving camera with low resolution. There is heavy glare from oncoming traffic.
![Page 42: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/42.jpg)
Dense Traffic Dataset
We introduce a novel dataset of 45 high resolution videos consisting
of dense, heterogeneous traffic.
We have carefully annotated the dataset following a strict protocol.
The videos are categorized by camera motion, camera viewpoint, time of the day, and difficulty level.
![Page 43: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/43.jpg)
Where to put your money in 2019 — it's not US stocks, according to Morgan Stanley (Emerging Economies)
https://www.cnbc.com/2018/11/26/stock-picks-morgan-stanley-upgrades-emerging-markets-downgrades-us.html
![Page 44: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/44.jpg)
Where to put your money in 2019 — it's not US stocks, according to Morgan Stanley (Emerging Economies)
![Page 45: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/45.jpg)
Dense
• Many agents (>3000) per Km of road length.
Key Ideas
Heterogeneous
• Different types of road agents present simultaneously, e.g., pedestrians, two-wheelers, three-wheelers, cars, buses, trucks etc.
Traffic Prediction
![Page 46: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/46.jpg)
Traffic Prediction
Combining multi-agent navigation, deep learning & dynamics
![Page 47: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/47.jpg)
Organization
48
• Pedestrian and Crowd Motions
• Heterogeneous multi-agent simulation
• Tracking urban traffic & Prediction
• Driver behavior modeling
![Page 48: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/48.jpg)
Modeling Driver Behaviors
• Most traffic accident happens are caused by dangerous reckless drivers
• “Aggressiveness” and “Reckless” are subjective metrics• Need navigation algorithms that can extract driving
behaviors from sensors/trajectories and perform safe navigation (Behavior-based Navigation)
![Page 49: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/49.jpg)
Aggressive Driving Behaviors
![Page 50: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/50.jpg)
If you are driving, which driver will you pay attention to?
Introduction NavigationFeature & Behaviors
TDBM
![Page 51: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/51.jpg)
Identifying Driving behavior allows autonomous systems to:
Pay extra “attention”
Avoid getting close to them
Re-run perception algorithms at higher resolution for those area
Introduction NavigationFeature & Behaviors
TDBM
![Page 52: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/52.jpg)
Main contributions
Feature extraction from trajectories in real-time
Trajectory to driver behavior mapping
Improved real-time navigation; Integrated with Autonovi-Sim
Introduction NavigationFeature & Behaviors
TDBM
[Cheung et al. 2018, CVPR; Cheung et al. 2018 IROS]
![Page 53: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/53.jpg)
Introduction NavigationFeature & Behaviors
TDBM
Outline•Outline of Algorithm
![Page 54: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/54.jpg)
Trajectory Database
Interstate highway-80 dataset• 45 minutes of trajectory• 1650 feet section of highway • captured with 7 cameras• tracked automatically with manual verification
Feature & BehaviorsTrajectory Database
User EvaluationFeature
Extraction
TDBM Navigation
![Page 55: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/55.jpg)
User Evaluation
Feature & BehaviorsTrajectory Database
User EvaluationFeature
Extraction
TDBM Navigation
![Page 56: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/56.jpg)
Feature Extraction
Feature & BehaviorsTrajectory Database
User EvaluationFeature
Extraction
TDBM Navigation
![Page 57: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/57.jpg)
Trajectory to Driver Behavior Mapping
TDBM Navigation
![Page 58: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/58.jpg)
Navigation improvements
Navigation
![Page 59: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/59.jpg)
Navigation improvements
Navigation
![Page 60: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/60.jpg)
Navigation improvements
Navigation
![Page 61: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/61.jpg)
Conclusions
•Behavior modeling using feature extraction
•Applied to highway traffic data
•Safe and improved navigation
Introduction NavigationFeature & Behaviors
TDBM
[Cheung et al. 2018, CVPR; Cheung et al. 2018 IROS]
![Page 62: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/62.jpg)
Crowd and Traffic Motion
• New algorithms for tracking pedestrians & traffic agents• Handle dense scenarios• Use models from social psychology for behavior modeling• Combine model-based and learning-based methods• Applications to crowd scene analysis and autonomous driving
63
![Page 63: Dinesh Manocha University of Maryland at College Park dm ...Dinesh Manocha University of Maryland at College Park dm@cs.umd.edu Prediction of Human Motion & Traffic Agents in Dense](https://reader034.vdocuments.mx/reader034/viewer/2022042317/5f06858f7e708231d41867e7/html5/thumbnails/63.jpg)
Acknowledgements
• Army Research Office• Baidu• DARPA• Intel• National Science Foundation
Questions: [email protected]
64