Target Tracking Based On Improved YOLO Algorithm In UAV Environment

With the rapid development of technology in recent years,UAVs are involved in more and more fields,and target detection and tracking algorithms for UAVs are gradually becoming the focus of research in computer vision.In the process of acquiring image information,UAVs are often constrained by light intensity factors,target deformation factors,target occlusion factors and scale change factors,resulting in low accuracy of target detection,discontinuous tracking,non-real-time tracking and even loss of targets.In order to achieve the goal of autonomous identification and real-time tracking of targets by UAVs,reduce the interference of UAV image acquisition by external factors and improve the detection accuracy and tracking performance of UAVs,this paper designs and builds a target tracking system in the UAV environment with image processing as the core technology in the field of computer vision.This paper proposes an improved target tracking method with YOLOv4 detection algorithm and ReID data fusion to correlate data between detection targets and candidate targets for UAV tracking of people.After YOLOv4 has identified and extracted features from the detection target,multiple selected candidate frames are filtered using HSV and LBP feature histograms to obtain the optimal candidate frame using non-great suppression,and after the target is identified,the target tracking method using YOLOv4 and ReID data fusion is then used to generate candidate frames for the target trajectory from Kalman filtering and the detected candidate frames The data is then correlated to determine the final tracking target.In particular,the Siamese network is used as a data correlator for the ReID backbone in Kalman’s trajectory tracking process,effectively solving the tracking drift phenomenon and applying it to long-time tracking.The process of target tracking by YOLOv4 and ReID data fusion algorithm is as follows: 1.The detector classifies and extracts the target’s bounding box,and uses the target detector and target tracker to classify and extract the target’s bounding box and capture the image information of the candidate box in the Kalman trajectory respectively;2.The object selector is used to select the candidate box in the trajectory to filter the initial optimal candidate box;3.ReID correlates the best candidate frame selected by the object selector with the tracked trajectory to confirm whether the trajectory information matches.The problems that affect the quality of the picture quality caused by various complex environmental factors during the process of collecting pictures are affected by various complex environmental factors,the pre-processing method for filtering and denoising with gaussian filtering is used for filtering and de-noising to reduce the impact of noise on the collected pictures.In addition to this,the grey-scale pre-processing method of weighted averaging is used to reduce the computational effort of the images.In particular,a BIMEF multi-exposure fusion image enhancement algorithm suitable for this environment is used in this paper to further improve the quality of the images in order to eliminate the effects of different lighting on the quality of the photographs.A vision system for UAV target recognition and target tracking was designed,using Pyqt5 to implement image enhancement,image greyscaling,image filtering,image correction,target detection and target tracking system functions.Several simulation tests and experimental tests were conducted on the system to realise the real-time tracking and display of targets by UAVs,and to verify the robustness and reliability of the UAV vision system.