Deep Point Cloud Semantic Segmentation And Its Application In Robotics

3D point cloud is a common geometric data type,which is widely used in computer vision,computer graphics and robotics.In recent years,the emergence of cheap 3D sensors has made the acquisition of point cloud data more convenient.At present,the interests in autonomous driving and intelligent robots in academia and industry also highlights the importance of point cloud pro-cessing and 3D scene understanding.Semantic segmentation and scene understanding based on point cloud,as well as its subsequent application in robotic perception,have become the current research hotspot.In this paper,the semantic segmentation of large-scale point cloud and the ap-plication of point cloud semantics in robotics such as scene recognition and SLAM(Simultaneous Localization And Mapping)are studied.The main work and innovative research achievements are as follows:1.Clustering-based Fast Semantic Segmentation of 3D Point Cloud.Aiming at the characteristics of LiDAR point cloud in outdoor large scenes,such as wide distribution range,sparsity and unstructured data type,a two-stage semantic segmentation framework of point cloud based on fast clustering and point cloud network is proposed.It overcomes the problems that traditional segmentation methods are difficult to obtain semantics and point cloud network is tricky to directly apply to the large-scale point cloud.It effectively combines the advantages of the two methods,reduces the amount of computing of the system,and processes them directly in 3D space,improving the segmentation accuracy and efficiency.By evaluating on public datasets,the proposed method has certain advantages compared with the advanced methods in the same period.2.Semantic Graph Based Place Recognition for 3D Point Clouds.Since existing methods based on points,lines and distribution features are difficult to achieve robust place recognition perfor-mance,and are difficult to be robust to point-view changes and partial occlusion,we propose a novel method based on semantic graph representation.The 3D scene is described at the seman-tic level by using the semantic information and semantic relations of point cloud,and the scene similarity is measured by the graph similarity network.By introducing the point cloud seman-tic information that is more robust to the scene environment changes,the algorithm improves the precision and recall rate,robustness and generalization performance,surpassing some of the most advanced methods on public datasets.3.Semantic-aided LiDAR SLAM(Simultaneous Localization and Mapping).In view of that there is no complete semantic LiDAR SLAM system currently,we introduce semantics into the existing LiDAR SLAM system.We enhance point cloud registration accuracy and robustness by semantic ICP and eliminate the cumulative odometry error by integrating semantic graph based place recognition method into the loop closure detection module.Our system can build a global consistent semantic point cloud map.Compared with the existing advanced LiDAR SLAM systems,our system achieves the state-of-the-art performance on public datasets.The proposed method lays a foundation for the following high-level tasks such as robot navigation,robot planning and human-computer interaction acting at the semantic level.