Design Of UAV Object Detection And Tracking System Based On Edge-cloud Collaboration

With the rapid development of artificial intelligence and unmanned aerial vehicle(UAV)technology,the research and application of artificial intelligence technology in the field of UAV is becoming more and more extensive.The UAV is rapidly applied to more and more industries because of its broad field of view,flexible movement,ability to operate in high-risk environments,and ignoring terrain restrictions.To solve the problem of detecting and tracking specific targets on the ground by UAV,this paper proposes an edgecloud collaborative object detection and tracking system based on the UAV experimental platform,and designs a distributed communication architecture for edge-cloud collaboration,which can achieve dynamic increase in the number of nodes and data interaction among nodes within the self-organizing network.Based on this,the edge-side UAV and the cloud server collaborate to accomplish the object detection task,as well as the continuous tracking of the specified target within the on-board computer of the edge-side UAV.Based on the realization of the above goals,the main research work of this paper is as follows:(1)The distribution of model structure internal parameters of the YOLOv4 object detection algorithm and the depth separable convolution of MobileNet were analyzed,and the feature extraction network of MobileNet depthwise separable convolution was proposed to replace the backbone network of YOLOv4.The common convolution in the PANet enhanced feature extraction network of the Neck part of the YOLOv4 model is replaced by the depthwise separable convolution to realize the lightweight of the model.In order to avoid the decline in accuracy caused by the reduction of model parameters,attention mechanism module is introduced into the replaced feature extraction network.The expression ability of the model is improved by strengthening key information and weakening useless information.The design of the Mobile-YOLONet lightweight object detection algorithm based on the Squeeze-and-Excitation is completed.The designed algorithm was trained and tested in the PASCAL VOC data set.Compared with the original YOLOv4 algorithm,the detection speed of the improved algorithm is increased by 8.4 FPS.Using the camera in the NVIDIA Jeston NX onboard computer for real-time object detection can achieve 13.2 FPS(Frames Per Second),and the accuracy of the detection category also meets the application requirements of object detection.(2)An edge-cloud collaborative distributed communication architecture based on ROS(Robot Operating System)was designed for the edge-cloud collaborative object detection algorithm,and the Mobile-YOLONet lightweight object detection algorithm based on the Squeeze-and-Excitation was deployed in the onboard computer of the UAV at the edge node.The YOLOv4 object detection algorithm is deployed on the cloud server to realize the realtime target detection of the ground target by the edge node UAV and the cloud server.By designing an edge-cloud cooperative selector based on Sugeno type fuzzy neural network,the parameters of the fuzzy control system such as membership function were adaptively learned and adjusted by BP(back propagation)neural network.Finally,it realized the enabling and disabling of the object detection algorithm on the cloud server in the edgecloud collaborative object detection,so as to improve the accuracy of object detection on complex image.(3)The object tracking and positioning algorithm of edge node UAV is studied,and a KCF tracking model algorithm with adaptive adjustment of multi-scale object frame is designed to realize the adaptive adjustment of object frame scale and solve the problem that the object frame in the original KCF algorithm cannot be adjusted,which leads to the interference information contained in the object frame.The coordinate conversion algorithm is designed to realize the coordinate information of the object in the pixel coordinate system through multiple coordinate transformations,and the GPS coordinates of the target object in the world geodetic system are obtained.The on-board computer of the edge UAV updates the information in real time to guide the UAV to follow the object at a fixed point.The effectiveness of the designed object tracking algorithm is verified by the UAV experimental platform.Finally,according to the experimental purpose and application requirements,the edge node UAV experimental platform and cloud server UAV ground control platform are designed,including hardware selection design and communication link design,software design and deployment,to realize the communication between edge node UAV and cloud server UAV ground control platform.Relying on this experimental platform,the experiment of object detection and tracking system based on edge-cloud collaboration has been successfully completed,and the stability of the experimental platform is verified.