Research On Vehicle Detection And Tracking Algorithm Based Video Streaming

Vehicle detection and tracking are the basic content and research hotspot of computer vision,and are widely used in video surveillance,autonomous driving,and intelligent transportation systems.With the development of deep learning,the object tracking algorithm based on Siamese Fully-Convolutional(Siamese FullyConvolutional,Siam FC)has attracted widespread attention due to its fast tracking speed,but the accuracy is low,and there is still room for improvement.In addition,vehicle detection algorithms based on machine vision have gradually become the research focus,and the pros and cons of detection algorithms directly affect the vehicle tracking effect.In order to improve the accuracy of vehicle detection and Siam FC algorithm to identify and locate targets,the main research work in this paper is as follows:(1)To address the problem of weak template image recognition in Siam FC,insufficient search image positioning ability,and poor feature extraction ability when the object appearance changes,this paper proposes a siamese dual attention mechanism fusion(Siamese Dual Attention Mechanism)based on Siam FC.Attention Mechanism Fusion,Siam DAMF)target tracking algorithm.First,Res2Net-50 is used as the backbone network for feature extraction,and multi-scale fusion of the grouping features of the residual module,which can deepen the network while improving the image recognition ability.Secondly,introduce the feature channel attention mechanism and the feature channel attention mechanism on the template image and the search image branch.Spatial attention mechanism,through hierarchical feature fusion and non-local pixel dependence,improve the discrimination ability and positioning accuracy of the object.Finally,to enrich the training data,use the potential relationship between the template image,the positive sample and the negative sample to design the Triplet Loss Function to improve the feature learning ability of the network.Experimental results show that on the two public data sets of UAV123 and VOT2018,the success rate of the proposed algorithm has increased by 3.7% and 5.1%,respectively,and the algorithm has reached 37 frames per second,meeting real-time requirements.(2)Aiming at the problems of low efficiency of traditional vehicle detection methods,high missed detection rate,and poor detection of small targets,this paper proposes an improved YOLOv3 detection algorithm based on Efficient Net YOLOv3-E.First,use K-means++ to cluster the training tags to determine the anchor box suitable for the vehicle width and height.Secondly,use Efficient Net with stronger feature extraction capabilities as the feature extraction network,and use 4 feature scales to fuse deep semantic information and shallow layers At the same time,it draws on the idea of residual block and the calculation method of depth separable convolution to improve the detection efficiency of small-scale vehicles.Finally,CIo U and Focal Loss functions are introduced to improve network convergence speed and detection accuracy.Experimental results show that on the UA-DETRAC and self-built data sets,the MAP,Recall and FPS of the proposed algorithm reach 90.9%,88.3% and 30f/s,respectively,which improves the detection accuracy of small target vehicles.(3)Aiming at the problems of inaccurate matching and inaccurate precision caused by the change of scale in vehicle tracking,this paper proposes a tracking algorithm based on data association.First,use the improved YOLOv3 vehicle detection algorithm to obtain the bounding box and feature information of vehicle detection.Secondly,use the Kalman filter to predict the tracking frame of the vehicle in the current frame,and use the Hungarian algorithm to associate the detection frame and tracking frame of the current frame.Finally use The smallest convex hull formed by multiple vehicle feature points counts the number of traffic flows.Experimental results show that the proposed algorithm can accurately trigger vehicle counting while maintaining the accuracy of vehicle tracking,with an average accuracy of 91.5%,and meets the application of real-time vehicle counting.