Vehicle Target Detection and Tracking in Complicated Scenes of Infrared Videos

Changke Chen, Mingkai Chen, Ping Yao

Abstract: Optical image automatic recognition is widely used in the fields of the disaster relief, public security maintenance, biological disaster prevention and medical treatment, etc. It is common to accomplish the recognition task with Hungarian Algorithm and Kalman Filter Method over objects detected by deep neural networks. However, this method has large errors under the conditions where there is camera movement, image rotation or target occlusion. This paper aims to detect the vehicle targets and obtain their coordinates from the infrared aerial video data with two stages: target detection and target tracking. In the target detection stage, YOLOv5 is used as the backbone network for model training. Then, in the target tracking stage, the following methods are designed and implemented based on the Hungarian Algorithm and Kalman Filter Method. First, data enhancement is designed to solve the problem of frame rotation and disturbance. Second, the appearance matching and variance verification methods are proposed to deal with the failure in prediction of target positions caused by camera movement. Third, the hidden frame temporary storage method is proposed to solve the occlusion problem, whose results are further confirmed by appearance comparison. Finally, in the output stage, the brightest point of five sample points is selected as the target point. With these methods, the missing detection and false alarm rates decrease, and the trace of the same target in the same video can be obtained even in complicated scenes. In the experiments, compared with the baseline algorithm, the missing detection rate of the proposed algorithm in the validation set is decreased by approximately 26%, and the track continuity score is improved by approximately 8.5%. Meanwhile, the size of the final prediction model is about 100MB while the results can be obtained within several seconds, which can meet the demands of real-time deployment of mobile platforms.