Research On Vehicle Detection And Tracking Algorithm Based On YOLOv4 From UAV Perspective

With the advent of the Internet and big data era,intelligent transportation systems will play an indispensable role in future smart cities,and vehicle detection and tracking is the most basic and critical part of it.Vehicle video is acquired by installing fixed cameras at heights,which is costly to maintain,and there are many monitoring blind spots in some remote areas,mountains,and tunnels.The UAV has a wide monitoring range and low cost of use,which can quickly complete the extraction of traffic vehicle information.Therefore,in terms of vehicle detection and tracking,the combination of drones and intelligent transportation systems can provide more convenience for intelligent transportation systems.Aiming at the problems of small vehicle size and complex environment from the perspective of UAV,and at the same time to meet the requirements of real-time algorithm,this paper proposes a vehicle detection and tracking algorithm based on traffic video from the perspective of UAV,and the main work and innovation points of the paper are as follows:(1)Design the overall framework of the vehicle detection and tracking algorithm in this paper.The network structure of the basic detection algorithm YOLOv4 and the tracking algorithm SiamRPN selected in this paper is analyzed and introduced in detail,and the vehicle detection and tracking algorithm is improved according to the characteristics of the UAV perspective in this paper,and the overall framework of the algorithm is constructed.(2)Design a lightweight MC-YOLOv4 vehicle detection algorithm.In order to improve the detection speed,a new backbone network CSPMobilenetv3 is designed,which combines the advantages of deeply separable convolution to achieve more efficient computing;In order to improve the accuracy of the detection algorithm,the improved k-means++ clustering algorithm is used to reacquire the size of the anchor.In order to make the network focus on the more important features of the target,the CBAM attention mechanism is added after the neck and backbone network of the network structure;Through the improvement of the confidence loss function and the category loss function,the problem of unbalanced number of vehicle categories is finally solved,and the generalization ability of the model is improved.Finally,a series of experiments are used to demonstrate the effectiveness of the algorithm.(3)Design a twin network vehicle tracking algorithm based on feature fusion.In order to avoid the need to manually mark the tracking box of the first frame in the video tracking task,the MC-YOLOv4 detection algorithm is used to automatically detect the vehicle in the first frame.In order to strengthen the feature extraction capability of the twin network and reduce the amount of model calculation,a new backbone feature extraction network Shuffle Netv2 is designed,and an improved lightweight residual network is used to achieve efficient feature extraction.In order to improve the robustness and accuracy of vehicle tracking,a feature fusion network based on the attention mechanism of Transformer is designed to establish the connection between global features and improve the interaction and information fusion ability between features.Finally,the effectiveness of the algorithm is proved by experiments.