Density Adaptive Registration of Large-Scale Point Clouds in Diverse Outdoor Environments

Point cloud registration is the foundation of collaborative multi-robot mapping tasks in outdoor environments. Due to the dynamic changes in communication bandwidth, the density of point clouds transmitted from the robot to the server will also change simultaneously, which will significantly affect the point cloud registration accuracy and even lead to the failure of collaborative mapping. To address this problem, we propose a density adaptive registration method for large-scale point clouds with varying densities in diverse outdoor environments. To extract robust point features and establish correspondences between point clouds to be registered, we use an improved MinkowskiUNet32 with a high-resolution point-based branch, which can provide high-resolution point information to supplement the coarse-grained voxel information. Then we propose an outlier rejection algorithm for point correspondences based on the relative height difference of the laser points between point clouds to be registered, which can eliminate wrong matches in the process of coarse registration. To adapt to point density variations in multi-robot outdoor mapping, a novel probability distribution-based point filter is presented to filter out points with dissimilar normal distribution, which can establish accurate point correspondences in fine-tuning. Extensive experiments on large-scale outdoor point cloud datasets KITTI, ETH, and Wild-Places demonstrate that the proposed method achieves state-of-the-art performance in accuracy and efficiency under the condition of varying densities of point clouds. In particular, our method achieves the best generalization on unseen domains between urban and field scenarios. Code is released at https://github.com/dutwjw/darls.