Real-time multiple object tracking based on optical flow
Hao Su, Yaran Chen, Shiwen Tong, Dongbin Zhao · 2019
Deep matching and Kalman filter-based multiple object tracking (DK-tracking) has been demonstrated to be promising. Traditional DK-tracking, however, relies heavily on high-performance detectors and assumes that the tracking target always moves at a uniform speed over short distances. But in complex environments, it is difficult to satisfy some special targets. So we propose a novel model called DK-Compensation-Flow-tracking (DKCF-tracking) which introduces a new inference pipeline to build data association for the state information of the missing targets across frames so that we can predict some simple results satisfying the conditions. Besides, we introduce optical flow information to achieve the target motion information and to guide Kalman filter prediction more accurately. Moreover, the accurate target positions predicted by Kalman filter will help to compensate for the missing objects. Experiments are performed on public datasets: MOT2016 and a competition datasets from The Chines Intelligent Vehicle Future Challenge. The proposed method achieves better performances compared to the DK-tracking with the assumption of a constant velocity movement.