CCTNet: A Circular Convolutional Transformer Network for LiDAR-based Place Recognition Handling Movable Objects Occlusion

Abstract--Place recognition is a fundamental task for robotic application, allowing robots to perform loop closure detection within simultaneous localization and mapping (SLAM), and achieve re-localization on prior maps. Current range imagebased networks use single-column convolution to maintain feature invariance to shifts in image columns caused by LiDAR viewpoint change. However, this raises the issues such as restricted receptive fields and excessive focus on local regions, degrading the performance of networks. To address the aforementioned issues, we propose a lightweight circular convolutional Transformer network denoted as CCTNet, which boosts performance by capturing structural information in point clouds and facilitating cross-dimensional interaction of spatial and channel information. Through extensive experiments on the KITTI and Ford Campus datasets, CCTNet surpasses comparable methods, achieving Recall@1 of 0.924 and 0.965, Results on the self-collected dataset further demonstrate the proposed method's potential for practical Hai Zhang is with the Centre for Composite Materials and Structures, Harbin Institute of Technology, Harbin 150001, P.R.China (e-mail: Materials and Structures, Harbin Institute of Technology, Harbin 150001, P.R.China (e-mail: juehundt@hit.edu.cn). Rhling et al. [14] proposed In this paper, a circular convolutional Transformer network a statistical-based method called Fast Histogram algorithm, with a regression loss is proposed for place recognition task which generates a one-dimensional histogram as a descriptor in scenarios with movable object occlusion. It treats the range image as Moreover, Scan Context [11] employed the polar coordinate a ring, utilizing multi-column convolution to learn local feature to map the point cloud into a two-dimensional (2D) matrix details, relationships between range image columns, and along radial and angular directions, serving as descriptors for circular structural features of the point clouds. However, crafting manual features usually a Range Transformer module is proposed to dynamically allocate requires domain-specific expertise, and manual descriptors weights to various channels and pixel regions, enabling exhibit limited robustness in handling variations and uncertainties the fusion and interaction of information from both channel in complex scenes [15].