Multiple Object Tracking Based on Tracking Compensation for Low-Resolution Scenarios

Abstract

Multiple object tracking is one of the critical directions in computer vision research. In the application of vision-based tracking methods, cameras are sometimes installed far from the targets to obtain a global view. There would be a large number of targets in the videos with relatively low resolution, which increases the difficulty of visual tracking. Applying existing tracking methods directly in such low-resolution scenarios will result in low recall and a large number of discontinued trajectory fragments, due to the instability of the target detection results. To alleviate the tracking performance degradation in low-resolution scenarios, a multiple object tracking method based on tracking compensation (MOT-TC) is proposed in this paper. A detector is applied to produce the candidate bounding boxes of the targets in the current frame. Then trajectories from previous frames are used to predict their states in the current frame. An assignment method is adopted to match the candidate bounding boxes to the predicted states. For the unmatched trajectories in the current frame, a single object tracking method for compensation is used to provide the target positions, which can increase the recall and reduce trajectory fragments. Meanwhile, a strategy based on the response map of single object tracking is designed to evaluate the tracking performance. Extensive experiments on low-resolution videos have shown that the proposed method outperforms the baseline and other state-of-the-art methods by a large margin.