LiDAR 기반 로봇 시스템에서 정확한 SLAM을 위한 Panoptic 스캔 컨텍스트 기반 루프 Closure 탐지 알고리즘

초록

Loop closure detection is crucial in a SLAM system to ensure a globally consistent map by reducing drift accumulation. Similar to other fields, handcrafted methods are starting to be replaced by learning-based approaches. Panoptic segmentation, which fuses instance and semantic segmentation, enables a richer understanding of the surrounding environment. The proposed method is based on the Scan Context++, further improved by integrating panoptic information into the generated descriptors. A panoptic similarity calculation, based on histogram and Earth Mover’s Distance, is performed and combined with the geometric distance similarity calculation, based on Scan Context++. Experimental results show improvements compared to the baseline approach and other methods, with a lower threshold suggesting higher sensitivity.

키워드