Markerless Articulated Human Body Tracking for Gait Analysis and Recognition

We present a particle swarm optimization (PSO) based system for markerless full body motion tracking. The fitness function is smoothed in an annealing scheme and then quantized. In this manner we extract a pool of candidate best particles. The swarm of particles selects a global best from such a pool of the particles to force the PSO the jump out of stagnation. Experiments on 4-camera datasets demonstrate the accuracy of our method on image sequences with walking persons. The system was evaluated using ground-truth data from a marker-based motion capture system by Vicon. We compared the joint motions and the distances between ankles, which were extracted using both systems. Thanks to the high precision of the markerless motion estimation, the curves illustrating the distances between ankles overlap considerably in almost all frames of the image sequences.