An improved video object tracking algorithm based on motion re-estimation

In this paper, we try to improve a video object-tracking algorithm using a motion re-estimation scheme. The usual object tracking algorithm is composed of two steps: object projection and boundary refinement. Since a single motion model cannot represent the movement of a semantic object, it is not precise enough to refine the object boundary. Therefore, we segment the video object by an efficient segmentation algorithm. If errors are large in some regions, they are selected for re-estimating motion models. By projecting those regions using new estimated motion models, we can obtain more reliable object projection, which can produce accurate tracking results. The simulation results show that the tracked boundary of the video object having complicated motion is improved