An efficient vehicle trajectory extraction method based on YOLO and improved DeepSORT algorithm

In order to speed up the extraction of high-precision vehicle trajectories from high-definition video, an efficient vehicle trajectory extraction method combining YOLOv5 with DeepSORT-CFLR algorithm, was proposed, which can efficiently obtianed the vehicle ID, time and location information. The feature extraction network of DeepSORT algorithm is substituted by FasterNet network to extract spatial features which can improve the target tracking speed. Learning rate adjustment strategy are imported to optimize the convergence effect of the model, which combined the learning rate warmup and decay method. The perspective transformation of the image keyframes is carried out, and the size of the vehicle target detection frame is unified to improve the accuracy of the vehicle trajectory data. The vehicle trajectory extraction results from high-definition video taken in Xi'an indicate that MOTA, MOTP and FPS of the DeepSORT-CFLR algorithm are 92.8%, 1.320 m and 12.157 frames/s, respectively. Compared with other classical algorithms, MOTA and FPS of the DeepSORT-CFLR algorithm are increased by 5.4%~10.5% and 2.739 frames/s~3.517 frames/s, respectively. The training time and computational volume of its feature extraction model is 4.5 h and 1.217 GFLOPs, respectively, which are both smaller than DeepSORT algorithm. While reducing the requirements of computer equipment, the speed and accuracy of vehicle trajectory extraction are improved.