SLGD-Loop: A Semantic Local and Global Descriptor-Based Loop Closure Detection for Long-Term Autonomy

In simultaneous localization and mapping (SLAM), the detection of a true loop closure benefits in relocalization and increased map accuracy. However, its performance is largely affected by variation in light conditions, viewpoints, seasons, and the presence of dynamic objects. Over the past few decades, efforts have been put forth to address these challenges, yet it remains an open problem. Focusing on the advantages of visual semantics to achieve human-like scene understanding, this research investigates semantics-aided visual loop closure detection methods and presents a novel coarse-to-fine loop closure detection method using semantic local and global descriptors (SLGD) for visual SLAM systems. The proposed method exploits low-level and high-level information in a given image thus combining the benefits of local visual features invariant to viewpoint and illumination changes, and global semantics extracted from the specific semantic regions. Robustness is achieved against long-term autonomy through the fusion of global semantic similarity with semantically salient local feature similarity. The proposed SLGD-Loop outperforms state-of-the-art loop closure detection methods on a range of challenging benchmark datasets with significantly improved Recall@N and higher recall rate at 100% precision.