OverlapMamba: A Shift State Space Model for LiDAR-Based Place Recognition
Jiehao Luo, Jintao Cheng, Qiuchi Xiang, Jin Wu, Rui Fan, Xieyuanli Chen, Xiaoyu Tang
AI summary
Problem
Existing LiDAR place recognition methods struggle to balance computational efficiency with robust yaw-invariant descriptor generation under viewpoint variations.
Approach
The method projects LiDAR point clouds into 1D range image sequences and processes them using a novel OverlapMamba block that combines bidirectional state space modeling with a stochastic SHIFT operation to efficiently capture spatial relationships and enforce rotational invariance.
Key results
- Lightweight network producing high-quality yaw-equivariant features
- Specialized OverlapMamba block maintaining linear computational complexity
- Bidirectional and SHIFT strategies enhancing yaw invariance and generalization
- Leading loop closure detection performance on KITTI, NCLT, and Ford Campus datasets
Why it matters
Provides a computationally efficient solution for real-time global localization and loop closure detection in resource-constrained autonomous navigation systems.
Abstract
Place recognition is the foundation for autonomous systems to achieve independent decision-making and secure operation. It is also crucial in tasks such as loop closure detection and global localization in Simultaneous Localization and Mapping (SLAM) technology. Existing LiDAR-based place recognition (LPR) methods use raw point cloud representations or multifarious point cloud representations as inputs, as well as employ convolutional neural networks or transformer archi- tectures. However, the recently proposed Mamba deep learning model combined with State Space Models (SSMs) has enor- mous potential in long sequence modeling. Therefore, we have developed a novel place recognition network OverlapMamba, which represents input range images as sequences. In a novel way, we use a stochastic reconstruction method to establish shifted state space models to compress the visual representa- tion. Extensive experiments on three public datasets demon- strate that OverlapMamba achieves competitive performance with real-time inference speed, which effectively detects loop closure even when traversing previously visited locations from different directions, indicating its strong place recognition ability and real-time efficiency. Our method has been implemented at http://github.com/SCNU-RISLAB/OverlapMamba.