Self-supervised 3D point cloud prediction for moving objects based on deep correlation information

Abstract In autonomous driving, the continuous movement of objects exacerbates point cloud sparsity, making future frame prediction a challenging task. While self-supervised methods excel in such dynamic environments by learning temporal evolution directly from unannotated data, existing approaches often model spatial structures and temporal dynamics independently, limiting their effectiveness in sparse and complex scenes. To overcome this limitation, we reformulate point cloud prediction as a unified spatio-temporal correlation learning problem and propose a novel framework, the Motion-guided Spatio-Temporal Point Cloud Prediction Network (MoCNet). The key idea is to use motion information to guide feature learning. Specifically, we design a dual-branch structure to separately capture spatial structure and temporal variation, and then integrate them into a unified representation. To better model long-range dependencies, we adopt a large-kernel convolution strategy to enhance feature interactions without introducing heavy computational cost. In addition, we employ a dual-frequency representation scheme that decomposes features into high-frequency and low-frequency components. This design helps preserve local details while maintaining global consistency. Experimental results show that the proposed method consistently outperforms existing approaches on multiple benchmark datasets. It achieves higher prediction accuracy and better generalization ability. Moreover, MoCNet supports real-time inference, with an average latency of 20 ms and 15 ms under different settings. This is significantly faster than the typical 3D LiDAR sampling rate. The predicted point clouds are also closer to ground truth in terms of shape, orientation, and spatial position, which improves reliability for downstream autonomous driving tasks.