Multi Target Detection And Tracking Algorithm Based On UAV Perspective

With the rapid development of UAVs,target detection and tracking based on UAV platforms has gradually become a major research hotspot in computer vision.In fact,UAVs have long been used in the military for area inspection and reconnaissance,and now there is also a demand for high-altitude area inspection in the civilian sector.However,in the existing algorithms,there is not a more stable method for dealing with high-altitude detection and tracking of multiple targets,which is often limited by complex scenes,environmental factors,information on the size of the detected targets,and so on.Therefore,this paper proposes a structure of UAV vision-based multi-target detection(car、bus、truck)and tracking only for the needs of civil area inspection,with the aim of effectively improving the detection and tracking capability of UAVs for small and dense multi-targets at high altitude and liberating labor,making it possible to effectively capture and extract target information in the area.The research work in this paper is as follows:1)Firstly,we analyze the characteristics for the data set consisting of acquisition and Visdrone,and carry out image pre-processing work according to the characteristics to enhance the image quality and reduce the difficulty of detection by weighted averaging,Gaussian filtering and restricted contrast adaptive histogram equalization.2)For the YOLOv5 target detection algorithm,clustering is performed to calculate 12 anchor frame sizes of 4 sizes,which are used for subsequent structural improvement and training,and CIOU＿LOSS is selected to replace GIOU＿LOSS and DIOU＿NMS is selected to replace NMS,which can solve the problem of lower degree of occlusion;in the network structure,the Backbone and In the network structure,the Backbone and Neck networks are improved respectively,both of which fuse the feature maps at 4 times downsampling,so the detection size is changed.The improved results show better results than the original algorithm,with a 4.4% and 1.4% increase in m AP,respectively,and the improved algorithm can detect targets that the original algorithm failed to detect.3)On Deepsort multi-target tracking algorithm,some adjustments are made to the detection-based tracking algorithm and the Reid network in the original algorithm is replaced with Fast Reid to make the performance better tracking effect.The improved results in its MOTA and MOTP increased by 5.1% and 3.8%,respectively,and its IDs decreased by 62.4)In the multi-target tracking application,the position of one front and one back tracking frame within the first 8 frames of the video is retained,the difference between the center pixel points of the two frames is calculated and converted to the actual distance corresponding to them,and the time of the 8 frames is also calculated so that the velocity can be calculated and the overspeed threshold is set,and the region of interest is intercepted for the overspeed target and the information of the target is retained to the maximum extent.