Research On Target Detection And Tracking Of Ground Vehicles From UAV Perspective

Unmanned aerial vehicle image acquisition and target detection and tracking of ground vehicles have important application prospects and research value in fields such as national defense and security,intelligent transportation,etc.Aimed at the characteristics of small target to be measured,special shooting angle,variable angle,etc.in the target detection of UAV to ground vehicles,the software and hardware platform of UAV was built,and the target detection and tracking algorithm of APV-YOLO combined with Deep SORT was proposed,which realized realtime detection and tracking of the edge computing platform at the UAV end,and achieved good performance.The main research contents are as follows.To achieve real-time target detection and tracking of different types of vehicle targets on the ground by intelligent drones,a software and hardware platform has been built.At the hardware level,a quadrotor UAV platform including flight control and airborne computers was built,and the hardware assembly of the UAV platform was completed.At the software level,the UAV flight control software design scheme is built at the flight control end,and the low altitude hover attitude control program is completed;On the Jetson Xavier NX edge computing platform,a software platform based on ROS is built,and the remote control software is used to control the terminal of the UAV airborne computer to achieve the corresponding functions.Finally,a drone simulation platform was built based on Gazebo,which improved the efficiency of drone development and verification.Aiming at the problem of insufficient real-time performance and accuracy of target detection algorithms,an innovative algorithm for the APV-YOLO network model is proposed.Specifically,the convolution layer is replaced by depth separable convolution to reduce the number of parameters and the amount of calculation,and the Swish activation function is used to improve the accuracy and stability of the model;Introducing CBAM attention mechanism to adaptively learn important regions in images;Using an algorithm that combines K-means and genetic algorithm to improve traditional clustering algorithms to avoid falling into local optima;CIo U Loss is introduced to optimize the loss function to improve the clustering effect,so as to improve the detection accuracy and efficiency of small target objects.Improved Model Processing 1920 ×The 1080 resolution low altitude vehicle video stream still guarantees a 3.16% performance improvement in average accuracy,reaching 7.33 FPS,even with a 19.5% increase in frame rate.An innovative multi-objective tracking algorithm based on APV-YOLO and Deep SORT is proposed to address the issue of target loss during vehicle tracking in drone video streams due to environmental occlusion.Using MobileNetV3 lightweight network to replace the target appearance network in Deep SORT while maintaining accuracy to further compress model parameters for extracting image features and integrating spatial color histograms.The optimized algorithm is deployed on the drone end.Through experimental comparison,the accuracy of vehicle target detection and tracking for multi target tracking has been improved by 16.6%,and the accuracy of multi target tracking has been improved by 7.4%.It can continuously and stably track vehicles,even if some vehicles are blocked by trees without ID exchange,which can meet practical application needs.This project provides a feasible solution for unmanned aerial vehicles to achieve autonomous flight and liberate manpower,and has important significance for the automated analysis of unmanned aerial vehicle images and the expansion of their applications.