OARegNet: Online Adaptive Point Cloud Registration for Embodied Intelligent Driving

Embodied intelligent driving involves autonomous systems interacting with and adapting to real-world environments through integrated sensing and cognitive processes, enhancing vehicle efficiency and adaptability. A crucial aspect of this is V2X technology, enabling vehicle-infrastructure collaboration and playing a key role in intelligent transportation. Multi-sensor data registration is essential for V2X data fusion and utilization. However, the initial registration accuracy of V2X point clouds is not satisfactory, with manual methods being time-consuming and inefficient. Current challenges in online V2X point cloud registration stem from differences in point cloud density, occlusions, and changing overlaps. Research in this area is limited. In this paper, OARegNet, a new online V2X point cloud registration network, is proposed to address the aforementioned challenges. Two core modules are introduced to deal with the issues. In the keypoint extraction layer, an adaptive density learning module is used to extract accurate keypoints from vehicles’ and infrastructures’ point clouds despite significant density differences and occlusions. Moreover, in the coarse registration network, an adaptive feature fusion module is introduced that utilizes a multi-head attention mechanism to learn rich point cloud information and estimate reliable point correspondences from point clouds with varying degrees of overlaps. Experiments are conducted on the DAIR-V2X dataset and a 95.19% recall rate is achieved, a 4.89% improvement over HRegNet, demonstrating the method’s effectiveness in online V2X point cloud registration.