Abstract:

This paper addresses the problem of object tracking by particle filter in catadioptric images. Whereas a large literature exists for traditional images, only a very few methods have been developed for catadioptric vision. We present two techniques to correctly deal with the strong distortion inherent to catadioptric images. First, we propose generating and diffusing the particles in the equivalent sphere space rather than in the original 2D catadioptric image. It allows to handle the specific distance associated to the distorted image in a general framework. Second, we use an adapted neighborhood to perform template matching by histogram comparison. It permits to compare templates quickly while using an active neighborhood for space-variant windows. An additional important feature of our system is that no interpolation is performed: we directly work on the original
catadioptric image. Experimental results demonstrated the validity of the proposed approach.