| LiDAR is widely used in Simultaneous Localization and Mapping(SLAM)technology for autonomous driving because of its high measurement accuracy,wide detection range and low illumination impact.Time of Flight(TOF)LiDAR can only detect the 3D coordinate information of a point cloud.Therefore,in dynamic scenarios,TOF laser SLAM algorithms are prone to introduce dynamic errors in point cloud alignment and motion residuals in mapping,which affects positioning accuracy and subsequent use of the map.In addition,TOF laser SLAM is prone to feature degradation when faced with a single feature scene,resulting in point cloud alignment failure.FMCW(Frequency Modulated Continuous Wave)LiDAR is a new type of sensor that can not only detect the 3D coordinate information of a point cloud,but also obtain the Doppler velocity information,which greatly improves the sensitivity to the surrounding environment.In this paper,with the advantage that FMCW LiDAR can capture Doppler information,a Doppler velocity-based LiDAR odometry and map building scheme is proposed to improve the robustness and accuracy of the point cloud alignment algorithm in dynamic scenes and single feature scenes,and to eliminate dynamic objects as much as possible to build a clean static point cloud map.The main work contributions are as follows.Firstly,the on-board FMCW LiDAR is constructed in the autonomous driving simulation simulator CARLA for generating dynamic scene FMCW LiDAR datasets,providing researchers with an inexpensive and convenient method for acquiring FMCW LiDAR data.Secondly,in order to eliminate dynamic errors in point cloud matching,a motion point cloud segmentation method based on random sampling consensus is proposed.The method extracts the motion seed points by fast and robust estimation of the LiDAR’s own velocity,and then performs region growing on the motion seed points to segment the complete motion object point cloud from the stationary background point cloud.Then,in order to cope with dynamic scenes and single feature scenes,the LPD-ICP algorithm is proposed.The algorithm improves the accuracy,robustness and real time performance of the matching algorithm through self velocity-based initial positional estimation,frame-to-local map matching,and Doppler velocity-assisted constraints to guide the direction of gradient descent for point-to-line and point-tosurface constraints.Finally,a forward point-level tracking and reverse grid elimination scheme is proposed in order to construct clean static point cloud maps.The scheme eliminates temporarily stationary objects that have not yet been registered in a single frame and cleans up misregistered motion point clouds in a local map. |
