Multiple views based human motion tracking in surveillance videos

Most work on activity recognition focuses on 2D image properties, holistic spatiotemporal representations, or space-time shapes in image domain rather than with 3D pose in a body-centric or world frame. Such techniques rely on advanced pattern recognition algorithms and interpreting complex behavioral patterns. In this work we posit that it is possible to achieve 3D pose tracking using videos recorded in multi-camera surveillance systems. We show experimental results that were obtained on PETS 2009 datasets. The estimation of the 3D articulated motion is achieved using a modified particle swarm optimization.