Tracking with Graph Cuts: Treating Clutter with Adaptive Penalties

Many techniques for tracking based on gradient descent cannot follow objects as they undergo large movements or deformation. On the other hand, multi-hypothesis trackers capable of handling such behavior are computationally expensive. The standard graph cut technique offers a middle ground, quickly capturing objects anywhere in the image; however, because of its global nature, it is prone to capturing outlying areas similar to the object of interest. This paper proposes a novel method to constrain the standard graph cut technique to regions of interest for tracking multiple interacting objects in near realtime. For each object, we introduce a penalty based upon distance from a region of interest. This results in a segmentation biased to this area. Also, we demonstrate the use of a track point filter for predicting the location of the object in each frame. The distance penalty is then centered at this location and adaptively scaled based on prediction confidence. We demonstrate tracking in grayscale and color videos.