Tracking under scaling and rotations using stochastic mean shift

We propose a method for tracking a single object in a moving camera scenario while the object undergoes scale changes and rotations. Mean shift trackers are well known for their performance in estimating object translation but, fail to perform under scaling and rotations. On the other hand, stochastic search based approaches are found to be superior in estimating the object positions under transformations. We propose to combine these two frameworks where the scaling and rotation parameters are optimized with respect to a Bhattacharyya coefficient based similarity measure. And, the object position is obtained by mean shift iterations with the scaled and rotated kernel support. Mean shift being a mode seeking algorithm, our approach thus realizes a stochastic gradient ascent formulation to solve the tracking problem. The proposed approach was found to have superior performance compared to basic mean shift and TLD when tested on the standard BoBoT dataset.