Tracking Multiple Indistinguishable Objects through Severe Occlusion

Occlusions are one of the biggest obstacles to reliable tracking. While many existing algorithms perform well while the objects of interest are separated, they often swap object identities when multiple objects occlude one another. We focus on the problem of tracking multiple, nonrigid targets through severe occlusions. Our occlusion reasoning algorithm takes advantage of reliable separated object tracking before and after occlusion events. We use BraMBLe to track separated objects and determine the starts and ends of occlusions. From the separated object tracks, our algorithm uses a simple depth-order heuristic to guess the identity correspondence in the start and end frames of the occlusion. We interpolate the correspondence to yield a guess of the objects’ per-frame motion. This motion guess is incorporated into layer-based affine optical flow estimation in the form of a prior probability. Our algorithm tracks the objects forward one frame at a time from the start of the occlusion and backward one frame at a time from the end of the occlusion. The motion guess is updated based on the current measurements of the object positions. These steps are repeated until the object tracks meet in the middle of the occlusion. We demonstrate results on challenging test video of three identical mice from a side view.