A hybrid framework for robust dynamic 3D point clouds removal

Abstract

When robot creates a map, dynamic objects can change the space and render the map unusable for navigation. Additionally, the vertical resolution of a VLP-16 LiDAR may be insufficient, making dynamic point removal challenging. To address these challenges, we propose a novel method for dynamic point detection and removal consisting of four components. Firstly, we introduce a multi-resolution heightmap to enhance the efficiency and precision of dynamic point recognition by segmenting ground points. Secondly, we address the issue of limited vertical resolution by fusing multiple scans to simulate additional scan lines and leveraging a multi-resolution range image for precise dynamic point elimination. Thirdly, we apply clustering and principal component analysis-based techniques to compute eigenvectors, facilitating the correction of misclassified static points. Lastly, we propose the utilization of a three-dimensional bounding box strategy to reinforce the monitoring of small static clusters with elevated probabilities of misclassification. These four components complement each other and are executed sequentially. We evaluated our method for both dynamic point removal and ground segmentation on the KITTI dataset and real-world environments. The results demonstrate that our method outperforms baseline methods and generates clean maps.