An Efficient and Accurate 3D SLAM Method for Dynamic Environment

Abstract

LiDAR-based 3D SLAM has always been one of the hotspots in the field of self-driving and mobile robots in the recent years. But how to build a robust and accurate map for a complex dynamic environment is still a challenging task, which attracts more and more attention in this community. In this paper, we proposed an efficient and accurate 3D SLAM method for complex dynamic environment, which mainly includes two stages. In the first moving object detection stage, an end-to-end full convolution semantic segmentation network (FCNN) is exploited to segment the potential moving objects accurately. Then the left point cloud is forward to the static SLAM module, which is based on direct point cloud registration method, the map is managed efficiently with the incremental kd-tree data structure. Additionally, an independent thread of the loop closure detection (LCD) based on the framework of multi-factor graph is adopted to further improve the accuracy and robustness of final outputs. With the elaborately design of the whole framework, the proposed method can work efficiently. The performance of the proposed method is validated with the benchmark dataset KITTI, the results show that by removing the dynamic objects, the stability and accuracy of SLAM can be greatly improved.