Research On Key Technologies Of Autonomous Local Obstacle Avoidance System For Unmanned Surface Vehicles

Unmanned surface vehicles,as the pioneers of future marine warfare in military use,are important tools for the state to safeguard maritime rights and interests.And in civilian use,it can effectively reduce the amount of manual work and the risk of operation,thus having a wide range of application requirements worldwide.As one of the core technologies of unmanned surface vehicles,local obstacle avoidance technology is not only the key measure of the intelligent level of unmanned surface vehicles,but also the premise that unmanned surface vehicles can successfully complete various tasks.By the title of “Research on Key Technologies of Autonomous Local Obstacle Avoidance System for Unmanned Surface Vehicles”,this paper analyzed the local obstacle avoidance requirements of unmanned surface vehicles and built the software and hardware framework of the local obstacle avoidance system.A vision perceptual method based on deep learning and motion planning method based on near-remote obstacle avoidance were proposed and experiments were conducted to verify its effectiveness.This paper provided a novel thinking and method for local obstacle avoidance application of unmanned surface vehicles which has high academic value and engineering application value.The main works of this paper are illustrated as follows:(1)By analyzing the relationship between global obstacle avoidance and local obstacle avoidance,the general steps of local obstacle avoidance problem were clarified.The characteristic constraints that need to be considered in the local obstacle avoidance problem of unmanned surface vehicles were studied.Based on the local obstacle avoidance requirement in unknown maritime environment and the selection of key components,the hardware framework and software framework of the local obstacle avoidance system composed of the shore-based information display system and the shipborne decision-making system were constructed.(2)The basic components of convolutional neural network were introduced in this paper.Combined with VGG16 convolutional neural network,the semantic segmentation model was used to classify the image into sky,water surface and land by pixel level,which realized the accurate extraction of the water boundary line.As for the requirements of target classification and target positioning in the local obstacle avoidance process,the Faster RCNN network structure combined with AlexNet convolutional neural network was constructed for obstacle recognition.Finally,the evaluation indexes were introduced to evaluate the performance of perception technology mentioned in this paper.(3)The model description and relative motion parameters of the unmanned surface vehicles and the encountering ships are derived.By introducing the ship collision risk model into the unmanned surface vehicles,the comprehensive collision risk index for unmanned surface vehicles based on fuzzy mathematics theory was proposed.Finally,based on geometry model and motion model description of obstacles,and combined with the dynamic characteristics of unmanned surface vehicles and International Regulations for Preventing Collisions at Sea,the trajectory re-planning method based on the improved velocity obstacle principle for remote collision avoidance was studied.(4)The vector field histogram was constructed by combining with the concept of certainty grid,and the environment information was processed into two levels of data compression,which divided the environment surrounded into feasible and infeasible areas.By introducing the A* heuristic search algorithm,the candidate motion of unmanned surface vehicles was optimized through forward prediction,and the short-range reactive obstacle avoidance method for unmanned surface vehicle was proposed.(5)The MODD dataset was introduced to evaluate the performance of visual perception method based on deep learning,the results show that the average accuracy,mean accuracy,mean IoU and weighted IoU of three categories that are divided by VGG16-based semantic segmentation network were more than 95%,and the mAP value of AlexNet-based Faster RCNN network can reach 72% with average frame rate 7.17 FPS.The V-REP semi-physical simulation platform was built to evaluate the performance of motion planning method,the results show that the algorithm can meet the COLREGS requirements while against static and dynamic obstacles in the scene under the preset global path,and can return to the original route after the collision avoidance until reaching the target point.The angular change trend is relatively smooth while remain the sufficient safety margin between the obstacles.Also,combined with the lidar front-rear frame fusion data,the VFH*-based short-range reactive obstacle avoidance method can obtain a reasonable local obstacle avoidance strategy which help avoid tail end collision or falling into a local minimum during the local obstacle avoidance process.Finally,the actual ship verification is carried out by constructing the unmanned surface vehicles platform and the experimental pool,the results showed that the visual perception technology based on deep learning can make up for the detection blind zone of lidar,and combined the short-range reactive obstacle avoidance method,the unmanned surface vehicles can reach the target point while effectively avoid the water surface obstacles both in single obstacles and obstacle groups environment.The obstacle avoidance trajectory is smooth with good obstacle avoidance effectiveness and target point accessibility.