An Unsupervised Learning Network for Large-scale LiDAR Point Clouds Registration

With the continuous development of autonomous driving technologies, the registration of outdoor large-scale LiDAR point clouds has become increasingly important. Unlike indoor small-scale object point clouds, large-scale point clouds have inherent sparsity, abundant outliers, and other limitations. These characteristics often lead to low alignment accuracy and high time consumption when applying existing methods to large-scale point cloud registration. To address these issues, we propose an improved point cloud keypoints extracting method based on rotation compensation and a convolutional end-to-end unsupervised point cloud registration network. The former enables reliable keypoints extraction. The latter further extracts global features from the keypoint point clouds obtained by the former method and learns the overlapping region information between the source and target point clouds using a spatial attention weight encoder, and it can be trained efficiently without pose ground truth. To ensure fast and effective convergence of the network, we introduce a chamfer distance loss based on dynamic overlap rates. We test our method on two outdoor large-scale LIDAR point cloud datasets: PandaSet and KITTI odometry datasets. The results demonstrate excellent and stable performance, when it is applied to either original consecutive frames or the case of simulating large angular variations in real-world scenarios between consecutive frames by randomly transforming the target frame. Moreover, by applying our method's registration results as initial values to the classic ICP, we not only achieve optimal accuracy and robustness but also significantly accelerate the convergence of ICP, enhancing the efficiency of precise registration.