Research and Implementation of Tire Tracking Algorithm Based on Multi-Feature Fusion
Panpan Yang, Huicheng Chen, Jiayun Yang, Yongming Yan, Lan Yao, Zhibin Zhao · 2021
In order to improve highway utility, the national government is vigorously promoting automatic toll at highway entrances. Cameras, as indispensable facilities, have been installed on sides of toll lanes. When vehicles gradually drive through the video area, wheels are detected through target detection and tracking technologies. The number of axles is accordingly detected and it is a factor mattering toll for large trucks. This is turned out an essential methodology for axle counting. Multiple Object Tracking(MOT) is the key point for it, while most existing algorithms fail for the proximity of tire positions and the similarity of tire appearance. For the challenge of proximate positions, a relative position relation model is proposed. It defines adjacent tires as a group according to their proximity. Furthermore, Kalman filter is introduced to calibrate preliminary results by reducing matching range. To conquer the challenge in reference to similar appearance, factors in terms of relative position, background, texture and color of grouped tires are involved. These factors are synthetically integrated with weighted convolution features to improve association in tires. Based on the real video data collected from toll highway stations, it is tested and evaluated that the proposed algorithm can effectively improve the accuracy of tire tracking in adjacent positions and similar appearance.