Detection And Tracking Technology For UAV Based On Binocular Vision

The rapid development of Unmanned Aerial Vehicle(UAV)technology has led to the continuous expansion of the application field of UAVs,while also posing certain threats,including flight safety,illegal intrusion,and theft of private information.Therefore,it is necessary to effectively monitor UAVs.To address the issue of low efficiency and high cost in traditional methods,this article combines binocular vision technology with deep learning based detection and tracking algorithms to design an UAV monitoring system that can detect and track multi-UAV,as well as achieve UAV ranging.The SF-YOLOv5 UAV detection algorithm is proposed to address the issue of insensitivity of target detection algorithms to small and multiple targets.Integrating swin transformer blocks into the backbone of the YOLOv5 detection network,the C3 Swin TR structure is proposed to better focus on global information;In the feature fusion stage,the fusion-concat method is used to assign different weights to feature layers of different resolutions to better balance feature information at different scales in the network;In the post-processing stage,DIo U-NMS is used for detection box screening,reducing missed and false detections by adding more filtering conditions.By evaluating on the TIB-NET dataset and the homemade dataset GUET-UAV,the m AP of the SF-YOLOv5 UAV detection algorithm can reach 60.92%,which is 3.3% higher than the original YOLOv5 algorithm,and the average processing speed can reach 58.34 frames/second.The SF-YOLOv5-DS UAV detection and tracking algorithm is proposed to address the common issues of missed detections and frequent ID switching in target tracking algorithms.This is achieved by combining the SF-YOLOv5 detection algorithm with the Deep SORT tracking algorithm.Using homemade UAV video data for evaluation,it was found that the tracking accuracy of SF-YOLOv5-DS improved by 10.27% compared to using the original YOLOv5 as the Deep SORT detector,reaching 92.59%,and the average processing speed was 37.81 frames/second.In order to meet the need for UAV samples during the system implementation process,a multi-UAV image dataset GUET-UAV containing 5174 UAV images was created for training and verifying detection algorithms.Collected 10 videos of multi-UAV for the implementation and comparison of tracking algorithms.Collected 6m-15 m static UAV binocular images and 15m-45 m dynamic UAV binocular images and videos for the implementation and error analysis of the entire system.An UAV ranging detection and tracking system is designed by combining binocular ranging method with SF-YOLOv5-DS algorithm.Firstly,the input UAV binocular image is passed through SF-YOLOv5 and the detected target information is output.At the same time,the input UAV binocular images are divided into left and right binocular images,and the depth map is calculated using the WLS-SGBM stereo matching method.Secondly,the detected output is used as the input of Deep SORT to achieve UAV tracking through prediction,matching,and updating.Finally,take the center point of the tracking box to provide feedback on the depth value of that point in the depth map,as the actual distance of the target UAV.The entire system can achieve ranging of UAV within a distance of6m-45 m,with a relative error within the range of 5%,enabling effective monitoring of UAV.