Robust LiDAR-Camera Alignment over Point Cloud Density Variation

For the first time, we raise and address the research problem of robust LiDAR-Camera alignment (LCA) over point cloud density variation, which is essentially important to the applications like autonomous driving and robot navigation. Our key idea is to generate the regular point cloud that holds the characteristics of density-insensitivity and evenness jointly for extracting robust and effective LCA feature. Our intuition lies in 2 main folders. That is, density-insensitivity aims to make the generated point cloud have a fixed number of points in local spaces. And, evenness requires that the points in local spaces can be evenly distributed to better recover the scene. The density-insensitivity and evenness ensure the generated point cloud can provide robust and effective information for LCA tasks. Specifi-cally, a 2D CNN model is used to achieve the density-insensitivity and evenness of generated regular point clouds. Additionally, a novel loss function is proposed to train our model without ground truth. The experiments on KITTI dataset verify superiority of our proposition. Our method can achieve comparable performance with state-of-the-arts LCA methods under original testing point clouds. As the point cloud density decreases, the performance of our method gradually outperforms them. When the point cloud density is very sparse, the performance advancement of our method is over 1.5 m / 0.4° on translation / rotation error with mis-alignment magnitudes of up to ±1.5m and ±20°.