Vehicle Tracking Based on Historical Intersection Over Union

Multi-object tracking (MOT) could be applied to many video analysis scenarios, such as vehicle speed estimation, vehicle re-identification, and vehicle abnormal behavior detection. A tracking task can be formulated as a data association problem, for which there exist many different types of solutions. Track-by-detection is one of the most common approaches for the MOT task. In this paradigm, the tracking algorithm relies on the detection results to decide whether detected vehicles in sequential frames belong to the same track. In our work, we developed a reliable vehicle tracker following this paradigm, while considering the balance between tracking efficiency and tracking performance. Our algorithm extends the existing intersection over union (IOU) tracker and improves upon it by fusing historical tracking information. In addition, our tracker allows tuning certain hyperparameters that lead to improved results, including the minimum confidence score, the maximum confidence score, the IOU threshold, and the length of a candidate track. We demonstrated the effectiveness and efficiency of our approach using the UA-DETRAC benchmark dataset. Our proposed approach runs at an average speed of 28 frames per second (fps), which is 16 faster than one of the baselines but 24 times slower than the other. With regard to effectiveness, however, our approach outperforms both baseline methods by more than 20% in most of the tracking performance metrics and achieves a 60% performance improvement in certain cases. We conclude that our tracker, which balances running speed and performance, could be useful for applications running in a real-time environment