Research On Detection And Obstacle Avoidance Of Underwater Fishing Nets Based On Vision

With the development of ocean value,underwater robots gradually replace man-made to become an important role in the field of ocean exploration.However,the existence of obstacles in the ocean has caused a great disturbance to the safe operation of underwater robots.Due to the scattered fishing nets in the ocean fishing or breeding operations,they are more difficult to detect than rigid obstacles.Therefore,this paper takes underwater fishing net obstacles as the detection target,and proposes an improved underwater fishing net detection method based on depth learning,aiming at the problems of current target detection technology,such as low detection accuracy of underwater fishing net,unable to adapt to its different morphological transformation and so on.In addition,an improved obstacle avoidance planning method based on deep reinforcement learning is proposed to solve the problems such as lack of generalization and low efficiency of exploration data.The specific research contents are as follows:First of all,this paper introduces the research status of deep learning target detection method and deep reinforcement learning obstacle avoidance algorithm,and analyzes the existing problems of the algorithm in depth.In view of these problems,the target detection method and obstacle avoidance planning method to be adopted in this paper are presented.Secondly,the underwater laser image acquisition system is constructed according to the underwater imaging characteristics,and the image enhancement is carried out to solve the problems of image color degradation and image blur.At the same time,the data expansion method is given.Thirdly,the paper analyzes the current mainstream convolution neural network,and puts forward the improved feature extraction network to solve the problem of its insufficient learning ability.And for the mainstream deep learning target detection methods can not meet the generalization performance of fishing net shape,a multi-scale prediction model of adaptive receptive field is proposed.In addition,according to the principle of binocular stereo vision,the classical stereo matching algorithm is compared and analyzed to realize the fishing net ranging and provide spatial information for obstacle avoidance planning.Finally,the advantages and disadvantages of mainstream obstacle avoidance algorithms are analyzed,and an improved obstacle avoidance planning algorithm is proposed to meet the requirements of intelligent agent generalization ability and effective data exploration in real obstacle avoidance planning scenarios.The training effect and efficiency of the improved obstacle avoidance algorithm model are verified by building a simulation environment.In this paper,the effectiveness of the improved deep-learning-based target detection method for improving the detection accuracy and speed of fishing nets and the effectiveness of the improved obstacle avoidance planning algorithm was verified through multiple comparison experiments,and the generalization of the improved algorithm was verified through the change of the simulation obstacle avoidance environment.Finally,this paper points out the existing problems and briefly explains the future development direction of deep learning target detection and deep reinforcement learning obstacle avoidance planning.