Detection And Location Of AUV Recovery Device Based On Deep Learning

With a greater voice and a greater role in polar affairs,China has more room to operate in areas such as exploration and development of polar resources,route development and scientific exploration.Therefore,China must have strong polar equipment to safeguard and fight for the legitimate rights and interests in the polar regions.With the development of various special-function unmanned submersibles in the polar regions and the problem of blind spots and large errors when positioning acoustic devices at short distances,the demand for underwater accurate docking devices is increasing.This topic mainly focuses on the optical detection and binocular positioning of the underwater docking recovery device by the polar AUV,and the targeted improvement of the YOLOv4 algorithm is carried out to meet the real-time requirements of underwater target detection while ensuring the accuracy.The binocular positioning method based on detection constraints has been proposed and verified by experiments.This paper first introduces the research background and research significance of the article,and conducts a comprehensive analysis of the research status of deep learning target detection algorithms,analyzes the advantages and disadvantages of each algorithm;expounds the advantages and disadvantages of the three types of sensors for underwater docking and recovery,and analyzes the underwater dual Comprehensive analysis of the current domestic and foreign research status of the target positioning method.Secondly,the YOLOv4 algorithm has been improved.To meet the real-time requirement of underwater robot docking and recovery,mobilenetv3 is used to make Yolov4 network lightweight.In order to solve the problem that it is more difficult to extract features from the complex underwater environment,the YOLOv4 network feature fusion ability is further improved,and the improved RFB structure is integrated into the YOLOv4 network.Aiming at the shortcomings of the poor accuracy of the YOLOv4 algorithm after weight reduction,the improved Bi-FPN network is integrated into the YOLOv4 network.The algorithm is validated on the public data set and the underwater docking device data set calibrated in this paper,and the improved YOLOv4 algorithm is compared with the current lightweight mainstream algorithms for ablation experiments.Then,analyze the underwater imaging model,underwater stereo matching constraints and Zhang’s calibration method.Considering the problem of light refraction due to the presence of air between the binocular camera and the waterproof sealing device,the analysis and deduction of the underwater calibrated imaging model with refraction parameters.Aiming at the low efficiency of traditional binocular vision positioning methods,a stereo matching method based on detection constraints is proposed.In the case of accurately detecting the position of the object in the left and right eye images,priority is given to binocular matching in the detection frame,and the weight outside the detection frame is reduced.This can reduce the sliding path of the stereo matching sliding window,thereby reducing algorithm time and improving efficiency.Finally,the algorithm of this paper is integrated into the underwater robot,and an underwater camera calibration experiment is designed to obtain the internal and external parameters and refraction coefficient of the binocular camera.On this basis,the localization experiment based on detection constraints is carried out,the mean error at different distances is analyzed,and the feasibility of this algorithm is verified.