Vision-based multi-person tracking by using MCMC-PF and RRF in office environments
Kanji Tanaka, Eiji Kondo · 2005
We propose a vision-based method for tracking multiple persons with gray-scale image sequence acquired by a monocular vision sensor in cluttered office environments. This method is based on a novel algorithm for acquiring depth of targets with primitive and robust features, position and size of targets. To cope with long-term occlusions caused by both fixed and moving objects, the method memorizes and utilizes history data of targets' state. We employ MCMC-based particle filter (MCMC-PF) to implement such domain knowledge including, interactions between targets, as well as radial reach filter (RRF) to extract objects in noisy gray-scale images. In experiments, the method could track multiple persons reliably, and recover from errors even when it loses sight of targets.