A Dynamic Object Removal Framework for LiDAR-based Environments

During the process of map creation, the presence of dynamic objects can disrupt the environment, rendering the map unsuitable for navigation. Additionally, the limited vertical resolution of a VLP-16 LiDAR sensor can present challenges in accurately identifying and eliminating dynamic points. To tackle these issues, we propose a dynamic point cloud removal method for pedestrians, which involves three essential components. Firstly, we adopt a novel ground point segmentation method to reduce the probability of misclassification of point clouds in subsequent processing steps. Secondly, we employ the k-means++ method to cluster each frame of point clouds, obtaining all potential clusters that may contain pedestrian point clouds. We deliberately refrain from discarding anything even remotely resembling a human. Subsequently, the multi-dimensional slice features and intensity attributes of the clustering results are extracted, and these are combined with the classification outcomes of the Support Vector Machine (SVM) to identify instances of pedestrians within the frame. Our method undergoes comprehensive evaluation in real-world environments, and the results demonstrate its superior performance compared to baseline methods.