A Translation-Tolerant Place Recognition Method by Viewpoint Unification

Place recognition serves as a fundamental component in tasks like loop closure detection and relocalization for mobile robots. Polar coordinate representations, such as Scan Context, which align with the data structure of range sensors, have become the most common data structure for point cloud descriptors in place recognition. While polar representations demonstrate the rotation invariance, they remain susceptible to translation variations. In this study, we introduce a novel approach: shifting the viewpoint of the original point cloud to construct the unified Scan Context, thereby mitigating translation variance. Our key concept focuses on identifying a stable, unified viewpoint for a given place and then pre-translating the point cloud accordingly. This naturally results in a descriptor devoid of translation variance. Importantly, within a given place, the viewpoint unification process tends to relocate the viewpoint to a similar position, irrespective of the original sensor perspective. In other words, the unified Scan Context becomes more closely associated with the place's structural characteristics than the physical location of the sensor. We validate our method through a series of comprehensive experiments encompassing synthetic scenarios and real-world datasets, showcasing its robustness in effectively handling translation variations.