Visual object tracking in 3D with color based particle filter

Abstract

This paper addresses the problem of determining the current 3D location of a moving object and robustly tracking it from a sequence of camera images. The approach presented here uses a particle lter and does not perform any explicit triangulation. Only the color of the object to be tracked is required, but not any precise motion model. The observation model we have developed avoids the color ltering of the entire image. That and the MonteCarlo techniques inside the particle lter provide real time performance. Experiments with two real cameras are presented and lessons learned are commented. The approach scales easily to more than two cameras and new sensor cues