Colour and Feature Based Multiple Object Tracking Under Heavy Occlusions

Tracking multiple objects in surveillance scenarios involves considerable difficulty because of occlusions. We report a composite tracker based on feature tracking and colour based tracking that demonstrates superior performance under high degrees of occlusion. Disjoint foreground blobs are extracted by using change masks obtained by combining an online-updated background model and flow information. The state of occlusion/isolation is identified by associating foreground blobs with object regions predicted using motion initialized mean-shift tracker (colour cue). The feature tracker is invoked in occluded situations to localize these with higher accuracy. We present results from dense traffic data with 5-15 objects in the scene at any instant. Overall tracking accuracy improves to 94.7% from 85.3% achieved by the colour only tracker.