Constraining Human Body Tracking.pdf

Constraining Human Body Tracking D. Demirdjian T. Ko T. Darrell Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, MA 02139 fdemirdji,tko,trevorg@ Abstract tracked parts. Within this linear manifold there will be other, non-linear, constraints, defined by joint angle limits Our paper addresses the problem of enforcing constraints and behavior patterns. Rather than attempt to specify these in human body tracking. A projection technique is derived algebraically we learn them from a set of joint angle training to impose kinematic constraints on independent multi-body data labelled with positive and negative examples of human motion: we show that for small motions the multi-body ar- pose. We then find a compact representation of the bound- ticulated motion space can be approximated by a linear ary of correct human pose using a support vector machine manifold estimated directly from the previous body pose. classifier. We propose a learning approach to model non-linear con- Using this framework we have developed a system that straints; we train a support vector classifier from motion can track pose in real-time using input from stereo cameras. capture data to model the boundary of the space of valid Motion of independent part is estimated using an ICP-based poses. Linear and non-linear body pose constraints are en- technique and an optimal articulated motion transformati