Research On Multi-target Detection And Localization Algorithm Of UAV Based On Computer Vision

Based on the constantly advanced technologies of artificial intelligence(AI)and unmanned aerial vehicle(UAV),the study on intellective UAVs has become one of the important orientations of the future high-tech and military fields.Multi-target detection and localization based on computer vision(CV)are the foundation for UAV to independently complete complex tasks,which can be applied to UAV inspection,UAV reconnaissance and other tasks.It has a very broad development prospect and application space.The thesis optimizes the existing target detection algorithm and target localization algorithm according to the characteristics of UAV application scenarios.The main research content is as follows:(1)Aiming at the problems such as small target size,dense target distribution and occlusion,wide image field and complex background,and limited hardware resources in the UAV perspective,the existing YOLOv5 target detection algorithm is optimized from three aspects,including network structure,loss function and label allocation strategy.In terms of network structure,SPD-Conv is introduced to solve the problem of information loss during convolutional downsampling,MSC3 module is introduced to improve the feature extraction capability of the original C3 module,and small target detection layer is introduced to improve the detection accuracy of small targets.The loss function uses SIo U function instead of CIo U function to improve the accuracy and convergence speed of the network.The OTA dynamic label allocation strategy is adopted to solve the problem of introducing harmful gradient in the label allocation of fuzzy anchor frame in intensive target detection,and improve the convergence speed and accuracy of the network.Finally,a UAV-YOLO target detection algorithm suitable for the UAV perspective was proposed,and experimental verification was carried out on the UAV aerial photography data set Vis Drone.Compared with the original YOLOv5 s,the m AP of UAV-YOLOs improved by4.8.Since accuracy of UAV-YOLO network is improved and the calculation amount is increased,UAV-YOLO network is compressed by using model pruning technology on the premise that the model accuracy is almost not reduced,and the pruning ratio is set at 20%to cut the UAV-YOLOs model.Compared with the original YOLOv5 s,the number of UAV-YOLOs parameters after cutting is reduced by 80.12 M,and the m AP is increased by 4.3 compared with the original YOLOv5 s.The improved UAV-YOLO in this thesis meets the requirements of small model and high precision,and provides a solution for target detection under the UAV perspective.(2)Aiming at the problems of low matching efficiency and low localization accuracy of traditional multi-target localization algorithm,a static multi-target localization algorithm of UAV based on multi-frame images is proposed,which can improve the accuracy of multi-target localization.Firstly,according to the algorithm,the object and the bounding box would be obtained through target detection algorithm from the input image.Then the multi-process multi-target tracking algorithm is used to complete the target matching between the multi-frame images.Finally,coordinate transformation and camera perspective projection model is used to establish equations in each process,and the least square method is used to solve the initial localization of the target.Gauss-Newton iterative method is used to optimize the localization results until the error converges or the target disappears in the image.Under the condition of a flight altitude of 200 m for unmanned aerial vehicles,experiments show that the average localization error of the UAV static multi-target localization algorithm based on multi-frame images can reach18.53 m,which is 23.26 m higher than the average localization accuracy of the traditional triangulation multi-target localization algorithm,and can be well applied to the static multi-target localization in multiple topography.(3)Aiming at the dynamic multi-target localization of UAV in flat terrain,a multitarget localization algorithm based on single frame image is proposed.Firstly,the target frame information is obtained by target detection in the input image.Secondly,the actual distance between the target point and the main point of the image is calculated by using the pixel distance between the target point and the main point of the image,and then the coordinates of the target point in the camera coordinate system are obtained.Finally,the longitude and latitude information of the target can be obtained by coordinate conversion.Under the condition of a flight altitude of 200 m for unmanned aerial vehicles,experiments show that in flat terrain areas,the average localization error of the UAV multi-target localization algorithm based on single frame image can reach 6.85 m,which is 21.03 m higher than the average localization accuracy of the dual-machine triangulation target localization algorithm,and can be well applied to the dynamic multi-target localization in flat terrain scenes.