Nonparametric data association for particle filter based multi-object tracking: application to multi-pedestrian tracking
Samuel Gidel, Christophe Blanc, Thierry Château, Paul Checchin, Laurent Trassoudaine · 2008
This article deals with the following issue: how to track a varying number of pedestrians through observations by means of a 4-plane laser sensor. In order to answer to the multiple target tracking problem and more specifically pedestrian tracking, we propose in this paper a statistical approach using a particle filter based on nonparametric data association methods. This approach allows to go beyond the conventional Gaussian assumption and to use as well as possible each particle during track/observation association by means of either a ldquoParzen Windowrdquo kernel method or a K-nearest neighbor algorithm. Simulated and experimental results show the relevance of this method compared to the usual Gaussian window methods.