Research On Ship Object Detection Method Based On LiDAR Point Cloud

Artificial intelligence systems with autonomous driving technology have broad application prospects in the road and maritime transportation fields,when intelligent systems perform environmental perception,Li DAR plays a vital role.Systems using active depth sensors can obtain accurate three-dimensional information of environmental scenes and are robust to lighting and weather conditions.The prerequisite for the intelligent system to achieve decision control is to effectively detect objects and obstacles in the scene,3D object detection based on Li DAR point cloud is a research hotspot in the field of deep learning.This paper studied the object detection method used in the Li DAR point cloud of road and sea large scenes,and achieved good detection results in the experiment.Firstly,we introduced the working principle of the 3D Li DAR sensor,analyzed the relationship between its resolution and distance detection accuracy;we also analyzed the characteristics of Li DAR point cloud data,and took the public KITTI dataset as an example to study the data annotation format of the point cloud dataset used for object detection task.Secondly,in view of the lack of available point cloud datasets for ship target detection in large marine scenes,we studied the use of ROS Gazebo simulation tools,referring to the design method of the KITTI dataset,and established a ship point cloud simulation dataset for marine scenes.We have completed physical simulation including sensor design,multi-category ship model addition,and marine scene simulation.We also realized the automatic labeling of data by writing script programs.By comparing the point cloud of the simulated scene and the measured scene,we analyzed the availability of the ship point cloud simulation dataset.In order to solve the problems caused by the sparseness and the limitation of information of Li DAR point clouds in large scenes,we designed a point cloud object detection method based on Hough voting network.We use a variety of data enhancement methods to enhance the training data scene,improve the convergence speed of network training.We used Point Net++,which is suitable for the original point cloud input,as the feature extraction backbone network.In further extraction of high-level abstraction point cloud features,we introduced the Hough voting mechanism,designed the VOTE module and its loss function,it can use seed points to generate virtual voting points that are closer to the center of the object in space,which improved the effect of feature extraction.We also analyzed the design of each component of our object detection network,and designed a loss function for the network.Finally,we conducted experimental study and result analysis on our point cloud object detection network.We designed different network specific structures and parameters,trained on the KITTI dataset and ship point cloud dataset,and evaluated the detection results of our model.The experimental results show that the accuracy of vehicle detection and multi-category ship detection reached 75% and 83% respectively,and both achieved good detection performance.