Online dynamic object removal for LiDAR-inertial SLAM via region-wise pseudo occupancy and two-stage scan-to-map optimization

Abstract

SLAM technology has become the core solution for mobile robots to achieve autonomous navigation. It provides the foundational information required for path planning. However, dynamic objects in the real world, such as moving vehicles, pedestrians, and temporarily constructed walls, affect the accuracy and stability of localization and mapping. Existing dynamic methods still face challenges, such as poor localization accuracy caused by reliance on IMU to provide initial poses before removing dynamic objects in highly dynamic environments, and decreased execution efficiency after incorporating complex additional processing modules. To improve positioning accuracy and efficiency in complex environments, this paper introduces dynamic object removal in front-end registration. Firstly, a two-stage scan-to-map optimization strategy is implemented to ensure the accuracy of poses before and after the removal of dynamic objects, where initial scan-to-map optimization is performed for precise pose estimation, followed by the identification and removal of dynamic objects, and a subsequent scan-to-map optimization to fine-tune the pose. Secondly, during the identification and filtering of dynamic objects, the method encodes the query frame and local map data that have already defined volume of interest (VOI) to generate a region-wise pseudo occupancy descriptor (R-POD), respectively. Subsequently, a scan ratio test (SRT) is conducted between query frame R-POD and the local map R-POD, identifying and filtering out dynamic objects region by region. This approach removes dynamic objects online and has demonstrated good mapping results and accuracy across multiple sequences in both the MulRan and UrbanLoco datasets, enhancing the performance of SLAM systems when dealing with dynamic environments.