Abstract:

We present an adaptive object tracking algorithm that is based on a novel consistency measurement computed recursively over a multifold of forward and backward frames. To obtain the multifold consistency score, the target object given for the current frame is searched in the consecutive frames forward in time without updating the object model. Then, a reinitialized model using the last observation is traced backward in time up to the current frame. Our hypothesis is that the object states before and after this process should be consistent for a successful tracking and a disagreement in states indicates a possible error. We utilize this score of each separate object to adjust the complexity of the corresponding core trackers, such as particle filters and mean-shift variants, and switch between these methods recurrently to extract the most reliable object tracks. Our results show the proposed recurrent tracking technique is capable of producing longer and more accurate trajectories, which is otherwise not possible for non-adaptive counterparts.