Correcting the tracker with memories

Visual tracking has attracted much research due to its importance in many practical applications such as video surveillance, vehicle navigation and augmented reality. Although existing trackers have made some success under various scenarios, long-term robust tracking in real-world scenes remains an open problem. Trackers are easily drifting away without external re-initialization due to drastic deformation, heavy occlusion and complex environments. We propose a framework to correct tracker by memorizing diverse states of the object. A memory pool of object state is constructed first, then maximum posterior probability is employed to estimate whether the object is present in the view and a best fit model is picked out from the pool for current frame. The qualitative and quantitative experimental results on various long video sequences demonstrate the superior performance of our method in comparison with other long-term trackers.