Research And Implementation Of Vehicle Target Tracking Algorithm Based On Deep Learning

Target tracking technology has great development potential in various fields,among which vehicle target tracking is one of the most important links in intelligent transportation,and has great research significance.For vehicle target tracking algorithm,because there are many factors affecting the accuracy of vehicle tracking in reality,the target tracking algorithm based on deep learning class is better,but it is often accompanied by high time complexity,large memory space occupation and high computing power demand,so that the algorithm is difficult to get out of the PC,which is not conducive to further practical application.At present,cost-effective embedded platforms emerge in an endless stream,and it is a general trend to achieve target tracking technology under the hardware environment of low power consumption.In order to separate vehicle target tracking algorithm from high performance hardware equipment and maintain good performance at the same time,the following researches are carried out in this dissertation:(1)Currently,TBD,a tracking strategy based on detection,is the main strategy used for vehicle target tracking.On this basis,combined with the characteristics of urban road traffic environment,the YOLOv5 s target detection algorithm and DeepSORT target tracking algorithm are deeply analyzed and studied in this dissertation.Aiming at the problems of large space consumption and slow detection speed caused by redundant calculation amount in YOLOv5 s,a series of improvements have been made.Firstly,Ghost Net is introduced to reduce the amount of network computation and accelerate the detection speed.Secondly,the attention mechanism of CBAM is integrated to improve the problem that is difficult to be accurately detected under various weather and illumination conditions.Finally,Soft-NMS is used instead of NMS algorithm to reduce the problem of missing detection caused by traffic congestion and other situations.Compared with experimental data on UA-DETRAC dataset,it can be seen that the improved YOLOv5 s algorithm ensures a higher average accuracy and significantly reduces the amount of computation,which verifies the effect of the improved algorithm.(2)The existing DeepSORT algorithm has the problem of heavy computation and low accuracy.Therefore,an improved YOLOv5 s is proposed to replace the original Faster R-CNN network to solve this problem.In the feature extraction part,this dissertation proposes to use the lightweight full-scale feature learning model OSNet to train the vehicle target weight recognition model,so as to enhance the feature extraction ability and improve the overall accuracy of the tracking algorithm.In order to solve the problem that the intersection ratio IOU calculation method cannot effectively judge the degree of coincidence between the detection frame and the predicted trajectory frame,the generalized intersection ratio GIOU distance measurement is used instead to improve the accuracy in the tracking process.(3)Transplantation of vehicle tracking algorithm was completed on NVIDIA Jetson TX2 device,and verification experiment was conducted.Experimental results show that the improved YOLOv5 s + DeepSORT algorithm can detect and track vehicles in road traffic scenarios better,and can also maintain good tracking effect on TX2 platform,which verifies the feasibility of applying the improved algorithm in this dissertation to embedded devices.