SemSCo: Semantic Frequency Domain Scan Context for LiDAR-Based Place Recognition.

Jiahao Li,Hui Qian,Xin Du
DOI: https://doi.org/10.1109/ICAR58858.2023.10406379
2023-01-01
Abstract:Place recognition plays a crucial role in robotic tasks such as relocalization and loop-closure detection. In this paper, we propose a novel semantic-based global descriptor. While global descriptors have shown promising performance in outdoor environments, they are sensitive to viewpoint variations and occlusions. To overcome these challenges, we exploit the semantic information derived from point cloud perception tasks to design a robust and distinctive global descriptor called Sem-SCo. Translation and rotation invariance is achieved through Fourier Transform and circular shift technique. Besides, due to the central symmetry property of the 2D Fourier Transform, we propose a fast method to eliminate the ambiguity of rotation angle and estimate the relative pose based on semantics. Exhaustive evaluations conducted across various scenes and datasets demonstrate that SemSCo outperforms state-of-the-art methods.
What problem does this paper attempt to address?