|
The on-board visual multi-object tracking task is a basic function of vehicle intelligent driving, which plays a connecting role in various applications and research fields such as traffic control, automatic driving and human-computer interaction. However, neither the method of using detection technology to locate and then using data association technology to generate target trajectory nor the method of combining detection and tracking fail to solve the problem that tracking performance (i.e. MOTA, MOTP, IDSW and other indicators) and tracking speed cannot coexist. This also makes it difficult for the existing multi-target tracking algorithms to be widely used in smart vehicles. To this end, we have designed an end-to-end deep learning multi-object tracking method that can be used in the vehicle, namely the self-query tracker, SQT. Specifically, the input of the algorithm consists of two parts :the current frame, and the tracking result of the previous frame. Firstly, the current frame is input to the backbone network to obtain the feature graph A. The feature graph A is input into the detection branch, and the detection object of the current frame can be quickly obtained through the regression between the heat map and the frame. Then the feature map A of the current frame is flattened and input to the coding-decoding network composed of Transformer. The tracking result of the previous frame is used as the query vector to obtain the position map of the tracking object of the previous frame in the current feature map. The final tracking result can be obtained by matching the two results (the detection result and the position mapping of the tracking object in the previous frame in the current feature graph). The training and verification based on MOT20 data set show that the inference time of each frame is about 44ms, and the multi-target tracking accuracy is 58.9%. The model is integrated into intelligent vehicle ROS platform for testing, and the test results show that the proposed algorithm can realize multi-target real-time tracking in complex traffic scenarios, and the algorithm has good practical application value. On the platform using RTX 2080Ti, the proposed method reached 15+ FPS and the MOTA score was 58.9 on the MOT20 dataset.
|
