Embedding ensemble tracking in a stochastic framework for robust object tracking

We propose an algorithm of embedding ensemble tracking in a stochastic framework to achieve robust tracking performance under partial occlusion,illumination changes,and abrupt motion.It operates on likelihood images generated by the ensemble method,and combines mean shift and particle filtering in a principled way,where a better proposal distribution is de-signed by first propagating particles via a motion model,and then running mean shift to move towards their local peaks in the likelihood image.An observation model in the particle filter incorporates global and local information within a region,and an adaptive motion model is adopted to depict the evolution of the object state.The algorithm needs fewer particles to manage the tracking task compared with the general particle filter,and recaptures the object quickly after occlusion occurs.Experiments on two image sequences demonstrate the effectiveness and robustness of the proposed algorithm.