Research And Implementation Of Obstacle Detection System Based On Semantic Segmentation For Unmanned Surface Vehicles

In recent years,China has strongly advocated the development of marine science and technology.As the automation of marine equipment,unmanned surface vessel is one of the main directions of the development of marine science and technology.In this paper,small unmanned surface vessel are taken as the research object,and water surface target detection is taken as the research direction.Combined with computer vision,we are committed to develop a high-precision algorithm for obstacle detection on water surface,which can be used in small unmanned surface vessel patrol,exploration,and sea rescue.In order to improve the detection accuracy of the algorithm,this paper starts with the preprocessing of water surface images and hope to eliminate the interference factors in the process of water surface obstacle detection.Through the practical application of detection algorithm,we find that,because the water surface is naturally smooth,water surface images at different time periods will include water surface highlight areas due to specular reflection.However,in many water surface target detection algorithms,the effect of specular reflection is mostly ignored,which leads to a reduction in detection accuracy.The traditional specular reflection removal method can only solve the specular reflection in a small range.For similar water scenes,when there are many pixels affected by the specular reflection,there will be incomplete specular reflection removal and even destroy the original color information of the image.Therefore,we first proposed a method for segmenting the water surface area and the non-water surface area,so as to limit the elimination of the highlight area to the water surface,and then proposed an improved intensity ratio model,for the pixels with the same diffuse reflection chromaticity,the property that the former is smaller the improved intensity ratio of pure diffuse pixels and highlight pixels can identify whether it is a highlight pixel.Besides,using this property can separate the specular reflection component contained in the highlight pixel,thereby achieving the purpose of removing specular reflection in a water surface image.On the other hand,in order to combine the method of removing the specular reflection in the water surface image with the water surface obstacle detection algorithm based on semantic segmentation,we deeply analyze the principle of water surface obstacle detection algorithm based on semantic segmentation.First,we introduce the traditional modeling methods using Markov random fields for image segmentation,and then analyze the advantages and improvements of the detection algorithm based on semantic segmentation.Next,we detail the modeling process and application of the probability map model for semantic segmentation of water surface images and the derivation process of the EM algorithm for parameter estimation of the model.To verify greatly the effectiveness of our proposed method for removing specular reflections from water surface images and to demonstrate the improved performance of the method combined with water surface obstacle detection algorithms,we collected and produced water surface reflection image data sets and water surface obstacle detection video data sets.At the same time,a comparative experiment is set up.The experimental results prove that removing the specular reflection in the water surface image greatly improves the water surface obstacle detection algorithm and makes the detection algorithm more robust.Finally,in order to improve the convenience of adjusting model parameters and obtaining experimental results,we designed and developed an experimental platform for surface obstacle detection system based on the researched methods.This system not only greatly improves work efficiency,but also facilitates man-machine interactive.