A novel loop closure detection algorithm based on crossroad scenes

Loop closure detection (LCD) is crucial for simultaneous localization and mapping (SLAM). Current LiDAR-based methods focus on global scenes and often overlook the rich geometric features of crossroads. These scenes, with their irregular contours, traffic facilities, and vehicles, provide valuable information that is not adequately captured by single-dimensional descriptors, leading to weak discriminative ability. To address this, a novel descriptor called singular value decomposition scan context (SVDSC) is proposed, leveraging singular value decomposition (SVD) to extract geometric features of crossroads, enhancing recognition capability. An adaptive weighted similarity calculation method is also introduced to improve accuracy by considering local feature values. Experiments on the Karlsruhe Institute of Technology and Toyota Technological Institute (KITTI), Jilin University (JLU), and self-collected crossroad datasets demonstrate the method’s superior performance in complex scenarios. Integrating this LCD algorithm into SLAM yields better mapping outcomes.