GICP-LOAM: Lidar Odometry and Mapping with Voxelized Generalized Iterative Closest Point

Abstract

The simultaneous localization and mapping (SLAM) have been widely applied for mobile robots in the GPS-denied environment. However, computational efficiency and localization accuracy is still a topic for further improvement of the SLAM system. Therefore, this paper proposes a SLAM framework by applying voxelized generalized iterative closest points (VGICP), which estimates voxel distributions by aggregating the distribution of each point in the voxel. As the voxelization approach can process the optimization in parallel, it can improve the efficiency of the calculation. Specifically, edge features and planar features are extracted for each frame in this SLAM system. Furthermore, the ground points and outliers of the feature point cloud are identified and removed. In the odometry module, the VGICP method is applied to make the feature points to feature points (key2key) registration, whereas feature points to map points (key2map) registration is used in the mapping module to achieve the mapping and localization functions. The proposed method is evaluated on the KITTI odometry benchmark. Experimental results show that this SLAM method has better accuracy and real-time performance over state-of-the-art open-source methods such as the A-LOAM, LeGO-LOAM, and hdl_graph_slam. Especially, the localization error is reduced from 19.269 to 5.955, and computational performance is improved by 10.39% compared to the A-LOAM.