When the robot or autonomous vehicle is driving, since the lidar will detect dynamic objects such as pedestrians and cars, these dynamic point clouds are stacked in the constructed map, which affects the accuracy of point cloud registration. Therefore, removing the dynamic point clouds of lidar has profound significance for improving the mapping and localization accuracy of the SLAM algorithm. Currently, most of the algorithms for removing dynamic point clouds are offline, which needs to compare the differences between the current frame and the map to determine the dynamic point cloud. These methods need to store many point clouds at historical moments, which require high computing and storage resources. As a result, it is impossible to determine the point cloud category in real time, and the removal efficiency of dynamic point clouds is affected. We propose a lidar SLAM algorithm that can remove dynamic point clouds online, which can both remove dynamic point clouds and improve the accuracy of lidar odometry.

Aiming at the problem that dynamic objects in the environment affect the mapping accuracy and efficiency, we propose a lidar SLAM algorithm DOR-LOAM to remove dynamic point clouds online. Firstly, the algorithm employs the current frame and the prior map to match the pseudo-occupancy area and divides the pseudo-occupancy area block by block for reducing the time consumption caused by the matching process. Secondly, it filters the bin regions that only contain ground point clouds, enhancing the robustness and speed of removing dynamic point clouds. Thirdly, it replaces the traditional method of utilizing whole-area matching and proposes a method of gridded pseudo-occupied areas. This enhances the recognition accuracy of the algorithm for areas containing both dynamic and static objects, and the recognition effect for small dynamic objects is particularly prominent. Fourthly, it adopts a sliding window method based on a dynamic removal rate to construct a priori map, greatly reducing the time consumption caused by matching dynamic point clouds. Meanwhile, the concept of dynamic point cloud memory area is proposed to further reduce the cost of repeated point cloud matching, making dynamic point cloud removal work in real-time.

The DOR-LOAM algorithm is evaluated on the KITTI dataset, and we assess the algorithm's point cloud removal accuracy and lidar odometry accuracy respectively. In terms of point cloud removal accuracy, the static point cloud preserve rate P_R, dynamic point cloud removal rate R_R, and F₁ score reach 94.13%, 97.11%, and 95.52% respectively. Compared with the latest open-source algorithm ERASOR, the P_R of DOR-LOAM improves by 4.00% (Table 1), and due to the utilization of gridded pseudo-occupied areas, the removal effect of small dynamic objects is better than other advanced algorithms (Fig. 6). In terms of lidar odometry, the relative translation error and relative rotation error are 0.81% and 0.0033 (°)·m^-1 respectively. Compared with the E-LOAM, T-LOAM, and NDT-LOAM algorithms, the relative translation error of DOR-LOAM is reduced by 0.75,0.12,and 0.09 percent point respectively (Table 2). Additionally, an ablation experiment is carried out on the algorithm, which verifies that the dynamic point cloud removal module can improve the lidar odometry accuracy (Table 3). In terms of the time consumption performance of the system, the algorithm greatly reduces the time consumption of point cloud matching through the point cloud memory area and the construction of sliding windows. The average time consumption per frame is 87.48 ms, which can meet real-time performance (Fig. 9).

The traditional lidar SLAM algorithm extracts point cloud line and surface features in different dynamic scenes, and the wrong extraction of non-static features will affect the point cloud matching accuracy. Thus, we propose a DOR-LOAM algorithm for the online removal of dynamic point clouds. Based on the spatiotemporal inconsistency of dynamic point clouds, the algorithm filters out non-correlated dynamic point clouds to eliminate its influence on lidar odometry at low cost and high efficiency and adapt to complex and changeable scenes. The verification results on the KITTI dataset show that the F₁ score of the algorithm for removing dynamic point clouds is 95.52%, and the relative translation error and relative rotation error are 0.81% and 0.0033 (°)·m^-1 respectively. The proposed algorithm not only yields better performance in the removal of dynamic point clouds but also improves the accuracy of the lidar odometry.