icme06 human action

Slide1: 

We propose a novel scheme to detect human actions in active video. Active videos are shot purposively, similar to the world seen from peoples’ eyes, and are object and action oriented, usually involving complex camera motions. We detect complex human actions in such videos. 1. Object Pre-localization with Composite Filtering 2. Finding object correspondence in successive frames This problem is formulated as an object labeling problem and solved by BP. Detecting Human Action in Active Video Hao Jiang, Ze-Nian Li and Mark S. Drew School of Computing Science, Simon Fraser University, Vancouver BC, Canada V5A 1S6 We propose a novel three-step human action detection scheme for active videos. Detects specific human actions by matching templates to video sequences, using a linear programming method. Successfully applied in general videos and TV hockey games. In future work we will study fusing other clues for action event detection, such as camera motions. Fig. 2. Object matching with composite filtering. For further information Please contact {hjiangb, li, mark}@cs.sfu.edu. Fig. 1. Detecting actions in videos. Finding action in a staged surveillance video: Finding actions in hockey games: Fig. 6. Finding action in an indoor active video. Template Action Detection Introduction Method Experimental Results Conclusion 3. Detail matching using linear programming In detail matching, we would like to find the point-to-point matching from template object to target object. The matching problem can be relaxed into a linear program: Comp- osite Temp- late Local valleys Template Target Object Matching Target-template Alignment, and Measure of Similarity Fig. 4. Object detail matching with linear programming. Fig. 7. Object pre-location and correspondence. Fig. 3. Object correspondence as a labeling problem. To improve the approximation, we use the following successive convexification scheme: For each site, set initial trust region to same size as entire target image Calculate matching costs for all possible candidate target points Find lower convex hull vertices in trust regions, and target point basis sets Build and solve LP relaxation Trust region small? Update control points Update trust regions No Yes Output results Delaunay Triangulation of feature points on template images Fig. 5. Diagram of successive convexification Fig. 8. Hockey player trajectories. Fig. 9. Finding shooting action. Fig. 10. Shortlist of hockey players for shooting action. Fig. 11. Finding another action in hockey games. Fig. 12. Shortlists of hockey players for action in Fig.11.