A Voxel-Based Multiview Point Cloud Refinement Method via Factor Graph Optimization

lidar enables fast reconstruction of the real world using high-precision point cloud maps. It usually requires the pose information (also called trajectory) of point clouds obtained by lidar at different times so that all scans are unified in the global coordinate system. Pairwise registration is one of the commonly used technology to estimate the poses of the lidar by measuring the relative pose of the scans and incrementally registering scans into the global map. However, the pose information obtained by pairwise registration largely suffers from low accuracy, inefficiency, and drift in trajectory. In this paper, we present a voxel-based multiview point cloud refinement method, which can effectively maintain the global consistency of point cloud maps. The proposed method store multiple point clouds into a voxel feature map indexed by the hash table, and then construct the factor graph using both point-to-point and point-to-plane factors to optimize the sensor's poses. A multi-scale voxel-grid optimization strategy is also presented to enhance the robustness as well as the accuracy of our algorithm. We conduct extensive experiments on the proposed method to demonstrate its accuracy and efficiency. It shows that the proposed method achieves higher accuracy than the state-of-the-art methods with a reasonable time cost.