Research On Underwater Target Detection And Tracking Method Based On Deeplearning

China has a vast sea area with rich resources,and vigorously developing Marine equipment is a necessary means to improve Chinaese Marine construction and governance capacity.This paper mainly focuses on the process of target detection and tracking in the underwater environment.In this paper,a lightweight target detection network based on YOLOv5 s is reaserched and the Deep SORT method is optimized.This paper firstly introduces the background and significance of the research,and comprehensively analyzes the research status of target detection methods and target tracking methods in China and abroad.The basic composition and function of convolutional neural network are introduced;The characteristics of Yolo series algorithms are analyzed and compared;The distance measurement related to target tracking and the Hungarian algorithm are introduced.For the real-time requirement of target detection method,the original YOLOv5 s network was modified with light weight.Firstly,the depth separable convolution method is used to lightweight YOLOv5 s network.Then,considering the possible detection quality degradation caused by the simplification of network structure,we choose to add CBAM attention module in YOLOv5 s lightweight network,and analyze the specific fusion mode of CBAM attention module in the network structure.The characteristics of various objects in the underwater data set are analyzed and compared.The diver data is collected in the experimental pool and the data set is made by ourselves.Finally,the effectiveness of YOLOv5 s lightweight network modification method was verified by experiments on self-made diver data.In order to solve the problem of identity jump caused by similar appearance of different track targets in multi-target tracking process,the calculation method of apparent feature similarity in the original method was optimized.Under the framework of detection-based target tracking method,YOLOv5 s lightweight network with CBAM attention module was used as detector,and the original Deep SORT method and the improved Deep SORT method were used as tracers to verify the effectiveness of the optimization method in MOT-16 dataset and self-built underwater target dataset respectively.