Scholarlink

With the development of Simultaneous Localization and Mapping (SLAM) technology, map-matching based localization can replace GNSS to provide location information for Autonomous vehicle. LiDAR-based Place Recognition (LPR) is one of the key components of map-based localization for self-driving cars in GNSS-free environments. Most of the existing efficient LPR methods for LPR use range images and single column convolution to ensure feature invariance for image column shifts caused by LiDAR viewpoint changes. However, the range image suffers from certain scale distortion, and the single column convolution has a restricted sensory field that cannot learn inter-column relationships, which reduces the network performance. In this paper, we construct the network SCMamba based on bird’s eye view, circular convolution and Mamba, expanding the receptive field by circular convolution while ensuring the image translation invariance, and designing the CirMambaAttention module for extracting serialized features to improve the accuracy. Our method is tested on four long time-span sequences of the NCLT dataset. The experimental results show that the average recall of our method outperforms the other five methods in all four sequences, which are AR@1 improved by 8.35–18.98%, AR@5 improved by 5.05–16.3%, AR@20 improved by 2.53–8.68%, showing that our method has some viewpoint variation and long-time span robustness.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

SCMamba: A LiDAR Based Place Recognition Network Based on Bird’s-Eye View and Mamba Architecture

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Long Yang,
  • Wuqi Wang,
  • Changlei Yan,
  • Chunyun Zheng,
  • Haigen Min

摘要

With the development of Simultaneous Localization and Mapping (SLAM) technology, map-matching based localization can replace GNSS to provide location information for Autonomous vehicle. LiDAR-based Place Recognition (LPR) is one of the key components of map-based localization for self-driving cars in GNSS-free environments. Most of the existing efficient LPR methods for LPR use range images and single column convolution to ensure feature invariance for image column shifts caused by LiDAR viewpoint changes. However, the range image suffers from certain scale distortion, and the single column convolution has a restricted sensory field that cannot learn inter-column relationships, which reduces the network performance. In this paper, we construct the network SCMamba based on bird’s eye view, circular convolution and Mamba, expanding the receptive field by circular convolution while ensuring the image translation invariance, and designing the CirMambaAttention module for extracting serialized features to improve the accuracy. Our method is tested on four long time-span sequences of the NCLT dataset. The experimental results show that the average recall of our method outperforms the other five methods in all four sequences, which are AR@1 improved by 8.35–18.98%, AR@5 improved by 5.05–16.3%, AR@20 improved by 2.53–8.68%, showing that our method has some viewpoint variation and long-time span robustness.