Research On Laser SLAM Based On Semantic Information

Simultaneous localization and mapping(SLAM)and its applications draw tremendous attention recently.Map construction is the foundation of autonomous driving.This thesis research on the construction of graph optimization-based outdoor map on an unmanned ground vehicle.It also explores the strategy to extract semantic information and construct semantic map.Our method also provides a solution of low-cost outdoor multi-layer map for park environment.The main contributions of this thesis are concluded in three parts.Firstly,the coordinates of sensors are aligned.The lidar and the combined inertial navigation measurement principle are established.Then time synchronization and extrinsics calibration between the two sensors are conducted.At last,the combined inertial navigation and lidar measurement are preprocessed for the use of date fusion.Secondly,autonomous navigation by use of a singular map is inaccurate.This thesis poses a low-cost multi-layer outdoor map made up by trajectory layer,static layer and dynamic layer.The trajectory layer is a GNSS topology map.The static layer contains point cloud map and grid map which both constructed by graph optimization method,while the dynamic layer mainly utilizes the information of current scan of lidar.The absolute position information of trajectory layer is used for global trajectory planning of unmanned ground vehicle.Then the scan-to-map matching between current lidar scan of dynamic layer and point cloud map of static layer provides location of the vehicle accurately.Finally,obstacles detection is implemented on the dynamic layer.Combining the grid map in static layer,it provides information for obstacle avoidance and local trajectory planning.Our method is evaluated in realistic environment and its feasibility is verified.Thirdly,to help with the lack of recognition of environment,this thesis adapts the method of semantic information extraction and mapping.The 3D point clouds from lidar are processed by a deep neural network.By projecting the point cloud to a depth image,the semantic information is extracted on that depth image.Then 3D semantic point clouds are obtained by projecting the 2D semantic information contained in depth image back.The KNN search method is used to eliminate the shadow presented in point cloud edges and finally the accurate semantic map is obtained.The semantic information helps with removing the effects of dynamic objects.Our method is verified in KITTI dataset.