| With the rapid development of society,traditional robots and mobile devices have been difficult to meet people’s requirements.Intelligent robots and self-driving cars that move autonomously have gradually become hot issues in human research.However,autonomous mobile devices need to have basic features such as autonomous perception,decision-making,and execution,and can assist and replace people in completing tasks in complex and dangerous scenarios,especially for some vulnerable groups.After years of development,the simultaneous localization and mapping technology(SLAM)of mobile robots has gradually developed and improved,and has achieved stable operation in some less demanding scenarios.However,the accuracy of the environmental map constructed by traditional laser positioning and mapping technology needs to be improved,lack of semantic information,and insufficient understanding of the surrounding environment to complete more advanced applications.With the rapid development of deep learning,building a specific neural network can semantically segment the point cloud information collected by lidar.Then it is integrated into the traditional point cloud map to realize the understanding of the semantic level of the surrounding environment by the mobile device,build a semantic map,and make up for the lack of semantic information and insufficient precision of the traditional map.Aiming at this problem,the main research of this thesis is as follows:(1)A Semantic Segmentation Method for 3D Laser Point Cloud Data.Using radar point cloud information as the carrier of environmental information,a fast asymmetric multi-branch residual neural network(AMBrnet)was researched and developed for semantic segmentation of Li DAR point cloud.Introduce multi-branch structure to extract feature information at the same time,and improve the ability of information encoding.After the network learning is over,the multi-branch structure is realized to be nondestructively fused to speed up the reasoning speed of the network.Therefore,the network has the characteristics of high-speed reasoning,light weight,easy deployment,and high precision.It has been tested and analyzed by the Semantic KITTI dataset,which proves to be more advantageous than the same type of network.The point cloud information will obtain a unique semantic label through the network,which provides semantic information reserve for the establishment of a large-scale environmental semantic map.(2)Research on semantic level environment map building method.Integrating the semantic segmentation network(AMBrnet)into the Su Ma mapping framework,and using the Su Ma++ back-end optimization method,the A-Su Ma++ algorithm,an outdoor large-scene semantic map establishment method that integrates point cloud semantic information,is obtained.Using the semantic information and vertex map segmented by the neural network as input,after the flooding algorithm,the corrected semantic label is output.Using semantic consistency to filter dynamic targets,and semantic ICP for point cloud registration to estimate pose information.The improved algorithm framework replaces the front-end part of the SUMA++ algorithm framework,and has better performance than Su Ma++ in building high-precision globally consistent semantic maps.Through the experimental comparison with SUAM++,the experimental analysis is completed on the KITTI Odometry Benchmark dataset,which proves that the A-Su Ma++framework has higher accuracy and better mapping effect than the SUMA++ framework before improvement. |
