Multiple-target tracking on mixed images with reflections and occlusions

Measurements arose from strong reflections combined with occlusions significantly degrade accuracy of multi-target tracking. Few methods have addressed this problem, and thus this paper proposes a robust multi-target tracker for mixed images with occlusions. For multi-cue integration using co-inference tracking, moving object detection significantly enhances motion cue based correction in the presence of reflections. Target templates are represented by sets of color and spatiality histograms. Joint likelihoods referring to both the target motion trajectory and appearance model of the co-inference fused state are computed. Thus each optimized particle weight with the criteria of maximum joint likelihood is more reliable in the face of reflections and inter-object occlusions. State estimation is achieved with the sample-based data association probability and occlusion confidence indicator. Experimental results show that the proposed tracker outperforms the-state-of-the-art multi-target trackers on images with strong reflections and inter-object occlusions.