SOLO-SMap: Semantic-Aided Online LiDAR Odometry and 3D Static Mapping for Dynamic Scenes

Abstract

Accurate and reliable online real-time localization and mapping are crucial for autonomous navigation of robot. Dynamic objects within the perception field can affect the accuracy of registration and localization, and also introduce ghost trail artifacts in the map, hindering robot planning and decision-making. While semantic segmentation can assist in perceiving object categories, it struggles to accurately segment moving objects. In this paper, we present SOLO-SMap, a real-time localization and static map construction framework based solely on LiDAR point cloud. We leverage semantic inference to identify potential dynamic points. And then, our instance-level true dynamic points removal is achieved by utilizing geometric rules based on moving point occlusion relationships and multi-object tracking (MOT) within a nearby temporal window in the pre-alignment stage. This design preserves stable static constraints while adhering to the static world model assumption of SLAM systems, benefiting accuracy and reducing drift, particularly in busy intersections. We evaluated the performance of SOLO-SMap in dynamic scenes on KITTI datasets and our self-made datasets, and conducted a comprehensive comparison with other methods, validating the effectiveness and robustness of the proposed method. A supplementary video can be accessed at https://www.youtube.com/watch?v=x-VKr3ag03M.