Research On Monocular Vision Positioning Method Of Underwater Vehicle Based On Deep Learning

Underwater monocular vision positioning is One of the most important key technologies in the visual perception of underwater vehicle.It has a very high application prospect in the fields of underwater security,mine detection,submarine terrain detection,submerged structure inspection and underwater vehicle docking.In this paper,a closed-loop monocular vision localization scheme based on deep learning is established on account of the requirements for underwater vehicle's monocular vision positioning task,and the underwater camera imaging model,object detection based on deep learning,monocular visual distance estimation and localization in the water,underwater image pre-segmentation based on target state estimation are also researched in this paper.Firstly,the influence of refraction on visual positioning in the process of underwater camera imaging is analyzed,an underwater camera imaging model is proposed on account of secondary projection,which divides the complex underwater imaging process into underwater nonlinear refraction sub-process and linear refraction sub-process in the air,and the model's iterative calculation method and model calibration method of the model are also researched.The equivalent focal length underwater camera imaging model is introduced to design the positioning contrast experiment.The results show that the accuracy and stability of the proposed model are better than the equivalent focal length model.Secondly,SSD300 is introduced to be an underwater object detector.A data augment method based on style transfer neural net is proposed on account of SSD300 training deployment difficulties.A contrast experiment is designed to evaluate the data augment method,the mAP(mean Average Precision)of SSD300 increased by 5.72%after data augment based on style transfer neural net,results also show that data augment method based on style transfer neural net can effectively reduce the cost and difficulty of data acquisition,improve the deployment of neural network training speed,high feasibility in underwater engineering application.Thirdly,the realization of geometric feature method and multi-frame method in this paper is researched on account of the target distance passive estimation problem in monocular visual positioning.The accuracy indexes of the two methods are established and the fusion estimation method of target's distance in monocular visual are studied.The experimental results show that the proposed distance fusion estimation method can effectively reduce the instability of one single method and improve the accuracy of target' s distance.Fourthly,in order to improve the detection ability of SSD300 for small objects,a method based on image pre-cutting is proposed.Target state estimation based on IMM-KF filter and image pre-cutting strategy are researched.In the relevant pool experiment,compared with DSSD algorithm and FSSD algorithm,and results show that the method based on image precutting improved the Precision of SSD300 to 91.4%,higher than that of DSSD(88.1%)and FSSD(87%),and the real-time performance was better.Finally,in order to verify the real positioning ability of our method in the underwater environment,the static underwater target positioning experiment and pool target positioning experiment are designed in the underwater environment,and the accuracy and stability of the algorithms are compared.