Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Sinisa Todorovic

263 views

Published on

From the Autonomous Systems @ OSU conference. http://research.oregonstate.edu/unmanned-systems-initiative

  • Be the first to comment

  • Be the first to like this

Sinisa Todorovic

  1. 1. Joint Inference of Groups, Events and Human Roles in Aerial Videos School of Electrical Engineering and Computer Science Sinisa Todorovic Autonomous Systems @ OSU Event June 30, 2015
  2. 2. Aerial Video Surveillance
  3. 3. Prior work: CAVIAR dataset Aerial Video Surveillance An example of our aerial videos
  4. 4. Our Problem Goal: • Groups • Events • Human Roles 1: Driver 2: Passenger
  5. 5. Our Approach Input Inference Output Exchange Box Input Inference Output Role AssignmentGrouping Event Recognition Guide Consultant Receiver Box Car Info Consult Exchange BoxGroup Tour Guide Consultant Visitor Deliverer Receiver Box Trajectories Frame registration Tourist Exchange Box Group Tour Info Consult Exchange Box Group Tour Info Consult
  6. 6. Noisy Input
  7. 7. Noisy Input
  8. 8. Inference Grouping
  9. 9. Pipeline: Inference Grouping Event Recognition Exchange Box Group Tour Info Consult
  10. 10. MCMC Inference Exchange Box Role AssignmentGrouping Event Recognition Exchange Box Group Tour Info Consult Exchange Box Group Tour Info Consult
  11. 11. Output Info Consult Exchange BoxGroup Tour Guide Consultant Visitor Deliverer Receiver Box Tourist
  12. 12. Challenges: Low Resolution
  13. 13. Challenges: Shadows, Top View
  14. 14. Challenges: Camera Motion & Partial Views X 10
  15. 15. Challenges: Structured Events
  16. 16. Exchange Box Play Frisbee Info Consult Pick Up Queue Vend Group Tour Throw Trash Sit on Table Picnic Serve Table Sell BBQ
  17. 17. Results Parse graph
  18. 18. Results • Baseline: Hierarchical clustering of trajectories + SVM • 3-fold cross validation Method Group Event Role Baseline 39.64% 16.94% 5.53% Ours w/o sub- event layer 40.41% 18.51% 8.69% Our full model 49.47% 32.84% 18.92%
  19. 19. Summary • New domain: low-res aerial videos • Unified representation and joint inference of groups, events and human roles • New mid-level feature: ST-templates • New aerial video dataset with detailed annotations of: – Human trajectories – Objects – Groups – Events – Human roles
  20. 20. Thank you!

×