The modeling and measurement of human movement are central to applications in visual surveillance and computer animation. The two most commonly studied approaches involve full 3D motion capture, on the one hand, and 2D tracking, on the other hand. Full 3D motion capture requires much higher resolution video than is typically available in surveillance. But modeling the types of everyday human movements of interest in surveillance - reaches, yanks, strikes, etc. - using dynamics of 2D silhouettes presents daunting computational challenges because of the high variability in the appearance of these movements. I describe an alternative approach based on the ballistic nature of common human movements. Our approach recognizes these ballistic movements independent of the movement's target-location and direction by modeling the ballistic dynamics. A video sequence is first segmented into ballistic subsequences without pose tracking, but instead based on global motion features of the body. The ballistic segments are then classified into strike and reach movements based on motion features. Each ballistic segment is further analyzed to compute qualitative labels for the movement's target-location and direction. Tests are presented with a set of reach and strike movement sequences. I will also describe the integration of this movement analysis system with appearance based object recognition, and show how the simultaneous recognition of movement and objects lead to higher classification rates for both.
Wednesday, February 20, 2008
Free and open to the public