Research On Underwater Obstacle Detection And Avoidance Based On Multi-beam Forward-looking Sonar

As the broadest part of the earth,the ocean possesses great development value.In recent years,people begin to exploit the resources and space of the ocean continuously,such as laying a large number of natural gas pipelines on the ocean floor for transportation.Therefore,the technology of using autonomous underwater vehicles(AUV)to detect whether a natural gas pipeline on the ocean floor is leaking or not is of great strategic significance.However,due to the complexity of the marine internal environment,autonomous underwater vehicles often have obstacles that interfere with navigation when performing tasks.Therefore,research on underwater obstacle detection and avoidance is particularly important.In this thesis,based on the data collected by the multi-beam forward-looking sonar mounted on an autonomous underwater vehicle,three aspects of obstacle detection method,obstacle avoidance method and command system are studied:(1)Based on the collected sonar image data,this thesis first proposes an obstacle detection method based on inter-class variance and small area suppression.Based on the fast calculation of the optimal threshold of the sonar image,this method aims at the situation where a certain number of noise areas still exist in the sonar image processing results.By continuously updating the optimal threshold,the detection results of obstacles in the sonar image are more accurate.(2)Based on the above method,this thesis proposes an obstacle detection method based on deep learning and improved threshold segmentation.The method uses the YOLO v3 network to detect the obstacle candidate areas in the sonar image.In the obtained obstacle candidate areas,an obstacle detection method based on inter-class variance and small area suppression is used to obtain the accurate obstacle areas.This thesis compares the proposed two methods with K-means clustering method,adaptive filtering threshold segmentation method,and Otsu threshold segmentation method on the two evaluation metrics of detection accuracy and operation speed.The experimental results show that the proposed obstacle detection method based on interclass variance and small area suppression has obvious advantages in two metrics compared with the three existing methods.The obstacle detection method based on deep learning and improved threshold segmentation proposed in this paper is based on this,which further improves the detection accuracy and processing speed.(3)On the basis of fully considering the actual working conditions and working environment of autonomous underwater vehicle,this thesis has designed a set of obstacle avoidance rules in a targeted manner,and combined with the obstacle detection results of sonar images,an obstacle avoidance method based on the contour of obstacle is designed to estimate a reasonable obstacle avoidance angle,which is transmitted to the master of the underwater vehicle to control the vehicle to avoid the obstacle.(4)This thesis designs a serial communication system and command system based on the RS-232 communication protocol,and realizes bidirectional communication between the obstacle avoidance board and the main control of the autonomous underwater vehicle.In this paper,the three modules of obstacle detection method,obstacle avoidance method and command system are finally combined and transplanted to a PC104 board as an autonomous underwater vehicle obstacle avoidance board card.In practical applications,the AUV equips with the obstacle avoidance board card successfully enters the water,and avoids obstacles during the voyage.