DR-REMOVER: An Efficient Dynamic Object Remover Using Dual-Resolution Occupancy Grids for Constructing Static Point Cloud Maps

Localization and navigation of vehicles using 3D point cloud maps is an important strategy in autonomous driving. In this work, we propose a new method for offline building static maps based on dual-resolution grids, called DR-REMOVER. Our method first extracts low-resolution grids containing dynamic objects, and then utilizes high-resolution grids for verifying dynamic points and reverting static points. The proposed method fully considers the sparsity of the point cloud and the environment features around the dynamic objects. As validated by experiments on SemanticKITTI and Apollo datasets, DR-REMOVER achieves more than 95% Preservation Rate (PR) for static points and Rejection Rate (RR) for dynamic points on all experimental sequences, which is significantly better than other State-of-the-Art methods. Furthermore, DR-REMOVER shows excellent performance on unmanned ground vehicle (UGV) dataset with highly crowded environments