Simultaneous Localization And Mapping Based On Point Cloud Semantic Segmentation

With the fast development of mobile robotics,applications such as self-driving cars and security robots have put forward higher requirements for the reliability of simultaneous localization and mapping(SLAM)systems,which requires unmanned systems to have a good ability to semantically understand their working scenes.For the tasks of SLAM and scene understanding,Li DAR is popular because of its high-precision ranging and running day and night,but the disorder of laser point cloud data leads to poor real-time performance of point-by-point semantic segmentation for large-scale point clouds.Inspired by the mainstream image semantic segmentation algorithms and based on the spatial dependence of the point clouds,our work can achieve laser point cloud semantic segmentation efficiently.Moreover,a semantic-enhanced scheme for laser SLAM is also proposed based on the output of point cloud semantic segmentation.To solve the problem of laser point cloud semantic segmentation,we propose a real-time framework of point cloud semantic segmentation based on a spherical projection model.Firstly,the mapping relationship between the 3-D laser point clouds and the pixels of the 2D range images is built,so as to extract point cloud features by using the proven backbone network for the image processing.Then the channel attention module is utilized to extract the dependency between each feature channel.While performing multi-scale feature fusion,the attention weight maps are calculated and assigned to corresponding feature maps with different fine-grained scales.Finally,the more precise semantic labels of 2-D range images are transferred to 3-D point clouds through the local context correlation and the neighborhood voting mechanism based on Euclidean distance.Based on the real-time semantic segmentation results of the above framework,a solution to integrate semantic information and SLAM algorithms is proposed in our thesis,which can effectively improve the accuracy of mapping and the reliability of semantic information.The specific optimization method of this scheme includes two aspects.First,we define the semantic similarity between point pairs,which is used to calculate the residual weight during the point cloud registration.This can reduce the influence of geometric mismatch and improve the accuracy of laser odometry.Second,a semantically consistent global map building method is proposed.According to the poses provided by odometry,the reliability of semantic information is improved by fusing multiple semantic observations.Based on reliable semantic tags,a place matching algorithm with rotation invariance and translation invariance is designed to build an effective loop closure detection method,which is more robust than the state-of-art methods.In this thesis,the KITTI dataset established by Karlsruhe Institute of Technology and Toyota America Research Institute is selected to test the accuracy and real-time performance of the proposed real-time semantic segmentation framework and verify that the integrated optimization scheme is beneficial to state-of-art SLAM frameworks.Moreover,the laser point cloud dataset built by our laboratory in our campus environment is also used for algorithm testing.The experimental results verify the effectiveness of the algorithm proposed in this thesis and its applicability to different data sources.