MORE: Multi-Frame Outdoor Point Cloud Registration with Temporal Memory Buffer

Jiuming Liu, Mengmeng Liu, Guangming Wang, Lihao Liu, Michael Ying Yang, Yunpeng Zhang, Francesco Nex, Hao Cheng, Hesheng Wang

Keywords: Point cloud registration, Temporal memory buffer, Dynamic History Weighting

TL;DR: We propose a novel outdoor point cloud registration paradigm enhanced by multi-frame temporal motion information, which improves the robustness to challenges like outliers and low-overlap inputs.

Abstract: Existing outdoor point cloud registration methods are commonly constrained by the pairwise input paradigm, which neglects sufficient temporal information intrinsically within consecutive LiDAR sequences. In this paper, we propose a novel Multi-frame Outdoor point cloud Registration network with tEmporal memory buffers (MORE). The key observation is that long-term temporal LiDAR sequences can provide rich global contextual information to complete sparse measurements, filter outliers, and address low-overlap problems, which further boosts registration performance. Specifically, two memory buffers are designed, including both the implicit memory feature buffer and explicit memory pose buffer, to store and dynamically update temporal pose-related features. We further leverage a Mamba-based temporal encoding layer to effectively integrate current motion features with history motion features. Moreover, a novel dynamic history weighting module is developed to adaptively rescale current and history pose-related features. Extensive experiments on three outdoor datasets demonstrate the superiority of MORE, surpassing all previous state-of-the-art methods by 32\% RTE and 17\% RRE reduction on KITTI, 37\% RTE and 4\% RRE reduction on nuScenes, and 29\% RTE and 9\% RRE reduction on Apollo-Southbay. Our method also generalizes well to the multiview indoor point cloud registration task with rather competitive performance on 3DMatch, 3DLoMatch, and ScanNet datasets. Codes will be released upon publication.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 1599