An ease-of-use stereo-based particle filter for tracking under occlusion

We describe a tracker that handles occlusion by clustering foreground pixels based on their disparity values. Stereo-matched foreground pixels, mapped from multiple views to a reference view, allow the tracker to recover occluded foreground regions in the reference view. The stereo algorithm utilizes a common plan view into which foreground regions from multiple views are projected and intersected to construct polygons that contain the ground plane locations of objects, followed by constraining the epipolar search to only those pixels with ground plane locations lying within these polygons. Consequently, a stereo-matched foreground pixel is easily mapped between views by first mapping its ground plane location using pre-computed homography, followed by intersecting the vertical axis passing through the mapped location with the epipolar line. Finally, tracking in a reference view allows a particle filter to be “seamlessly” integrated, so that uncertainties can be effectively dealt with. Experimental results illustrate the effectiveness of our algorithm.