Multi-object Tracking Combines Motion and Visual Information

Real-time online multi-object tracking is a fundamental task in video analysis applications. A major challenge in the tracking by detection paradigm is how to deal with missing detections. Visual single object trackers (SOTs) have been introduced to make up for the poor detectors in the assumption of appearance continuity of tracking objects. However, visual SOTs may easily be confused by the invaded foreground when occlusion occurs. In this paper, we propose to combine object motion information and appearance feature to improve the performance of object tracker. We use a lightweight re-identification feature to monitor occlusion. A Kalman filter, as the motion predictor, and a visual SOT, as the appearance model are worked together to estimate the new position of the occluded object. Experimental evaluation on MOT17 dataset shows that our online tracker reduces the number of ID switches by 26.5% and improves MOTA by 1-2% compared to the base intersection-over-union (IOU) tracker. The effectiveness of our method is also verified on MOT16 datasets. At the same time, the tracking speed can reach 29.4 fps which can basically achieve real-time tracking requirement while ensuring accuracy.