Research On Detection And Recognition Method Of USV Surface Target Based On Lidar

With the fierce development of unmanned driving,the unmanned surface vehicle(USV)as an unmanned carrier on the water has gradually entered the public’s field of vision.The selection of unmanned boat sensing sensors mainly includes Li DAR,vision,nautical radar and sonar.Due to the advantages of precise ranging and strong anti-interference ability,Li DAR is often used in short-distance environment perception of USV.However,the currently published algorithms cannot meet the task of detecting and identifying obstacles in real water scenes.Therefore,as the USV equipped with Hesai Pandar40 Li DAR to be the research object,the detection and recognition methods of water surface obstacles under the real water surface scenes is studied to support autonomous navigation and environmental perception of the USV.We takes the scanning scene of the water surface environment as the background,and divide the research into three parts: point cloud preprocessing,target detection and target recognition.The research content of this article is as follows:For the point cloud preprocessing,the relevant indicators and the scanning topology model are briefly introduced,Meanwhile,the horizontal and vertical point cloud resolution characteristics of the Li DAR is analyzed in detail.Aiming at the black and white points noise in the original point cloud of lidar in the water scene,a KNN numerical statistical filtering algorithm based on distance normalization is proposed to filter the white points,and a four-point collinear criterion is proposed to quick judgment and completion of the black points.Experiments prove that the above algorithm can effectively filter the black and white points noise in the scene and improve the quality of the scene point cloud.Aiming at target detection,the characteristics of point clouds in water surface scenes are analyzed in detail,which explains the reasons why it is difficult for existing target detection algorithms to meet target detection in water surface environment from the point cloud organization scheme.Based on the above difficulties,the front-view projection grid is first introduces as a point cloud organization structure to provide structural support for the algorithm design in segmentation accuracy and time efficiency;For wave filtering,an improved RANSAC algorithm combined with the point cloud vertical attributes is proposed,which avoids the damage to the point cloud of obstacles during the wave filtering process.For target clustering,a clustering criterion based on the angle threshold of adjacent points is proposed,and the fast clustering algorithm of obstacle point clouds is completed by using the fast marking algorithm of connected domains on the front-view projection grid map,which improves the clustering effect with high efficiency.Experiments show that the algorithm proposed in this paper has high segmentation performance and extremely low time complexity,and can effectively complete the target detection task of USVs and provide support for subsequent target recognition tasks.Aiming at target recognition,a typical water surface target point cloud recognition data set is firstly proposed in this paper.Based on that,the Light-Point Net and Multi-Scale LPoint Net are proposed to solve the target recognition problem in the marine environment.The experiment proves that the water surface point cloud recognition data set produced in this paper provides the premise for the deep learning to introduce the water surface point cloud target recognition task.At the same time,the two recognition network structures proposed in this paper can effectively complete the water surface target recognition task,compared with the traditional manual descriptors.The method in this paper has obvious advantages in terms of recognition effect and time efficiency,which can better support the unmanned boat’s perception of the environment.