Research On Vehicle Detection And Tracking Algorithm Based On Uav Traffic Video

Most vehicle detection and tracking are based on traditional fixed surveillance cameras.Fixed cameras have some disadvantages such as poor flexibility and limited field of view.UAV is flexible,has a wide monitoring range and low target occlusion.In a special period of time,the area without fixed cameras can be monitored,which effectively makes up for the shortage of fixed and limited monitoring area.Therefore,the vehicle detection and tracking algorithm based on the road traffic video captured by UAV is a worthy research and has a high practical value.Due to the traffic video from the drone’s perspective has problems such as small target size,and the tracking speed should also meet the real-time requirements.For this reason,a real-time vehicle detection and tracking algorithm for UAV traffic video was proposed in this paper.The main work and innovation are summarized as follows:(1)The general target detection algorithm was studied and analyzed.Firstly,the principle and structure of convolutional neural network were analyzed.Secondly,the principle of target detection algorithm based on two-stage and one-stage were studied.By comparing the detection accuracy and speed of the two types of algorithm,the YOLOv3 algorithm,which is very effective for the detection of dense small targets,was selected as the basic detection algorithm finally.(2)Due to the small size of the target in the traffic video captured by UAV,the accuracy and speed of the existing target detection algorithms is difficult to meet the requirements.In this paper,a vehicle detection algorithm YOMOv3-CIo U based on improved YOLOv3 was proposed.The k-means++ clustering algorithm was used to obtain an anchor box suitable for the dataset,the Mobile Netv3 network was used as the feature extraction network to realize network structure lightweight,and the CIo U Loss was used as the regression loss function to train and test the algorithm on the vehicle dataset.The improved algorithm was reduced the amount of parameters to two-fifths of the original,and the speed is increased by 10 frames per second.(3)Aiming at the problem of ID switch in multi-target tracking,a multi-vehicle tracking algorithm based on metric learning was proposed.YOMOv3-CIo U was used as the vehicle detector,and Kalman filter was used to predict the trajectory of the previous frame.The deep feature extracted based on deep metric learning was used as the appearance feature of the vehicle.The linear weighting of the motion information similarity based on Mahalanobis distance and the appearance feature similarity based on minimum cosine distance was used as the comprehensive similarity matrix.Finally,the Hungarian algorithm was used to solve the comprehensive similarity matrix,and the data association between the two frames was carried out to get the final tracking results.The experimental results showed that the proposed algorithm was achieved a faster tracking speed,reaching 17 frames per second without significantly reducing the tracking accuracy.