Tracking multiple objects in a scenario that exhibits complex interaction is very challenging. In this work, we propose a multi-resolution framework for multi-object tracking in complex wavelet domain to resolve the challenges resulting from occlusions and splits. In the proposed approach the appearance model is computed at high resolution to achieve more discriminative power to object model, whereas all other tasks are performed at low resolution to get benefited from noise resilience nature of wavelet domain. In this paper, we have also discussed a simple and effective unimodal background subtraction approach to extract moving objects by exploiting the low sub-band characteristics of the object image in complex wavelet domain at low resolution. A scheme exploiting the spatial and appearance information is used to detect and correct the occlusion and split state. Experimental results illustrate the effectiveness and robustness of the proposed framework in ambiguous situations in several indoor and outdoor video sequences.
Read full abstract