An Online Dynamic Point Separation and Removal SLAM Frameworks for Dynamic Environments

Dynamic objects in the environment can compromise map quality and, in severe cases, lead to robot localization failures. To address this issue, this paper proposes a simultaneous localization and mapping (SLAM) framework with dynamic point removal capabilities, which incrementally filters out dynamic points during the mapping process to enhance map accuracy and localization reliability. The framework consists of two main modules: the SLAM module and the dynamic point removal module. The SLAM module, based on Fast-LIO, incorporates novel loop detection and filtering algorithms to improve long-term mapping accuracy, while the dynamic point removal module optimizes the map by eliminating dynamic points. The dynamic point removal module employs three key methods. Firstly, to enhance dynamic point identification accuracy and minimize misclassification, a novel multi-resolution height map method is introduced. This method effectively segments static ground points and directly preserves them as static points. Secondly, a visibility-based approach is employed to maximize the removal of suspected dynamic points by comparing range differences between the local map and the current frame. Finally, K-nearest neighbors and principal component analysis methods are utilized to compare feature vectors between clusters, facilitating the recovery of static points that may have been erroneously removed. The proposed method is validated via public datasets and real-world scenarios, demonstrating significant improvements in dynamic point recognition as well as in localization and mapping accuracy compared to other state-of-the-art methods.