Hierarchical Loop Closure Detection for Long-term Visual SLAM with Semantic-Geometric Descriptors

Modern visual Simultaneous Localization and Mapping (SLAM) systems rely on loop closure detection methods for correcting drifts in maps and poses. Existing loop closure detection methods mainly employ conventional feature descriptors to create vocabulary for describing places using bag-of-words (BOW). Such methods do not perform well in long-term SLAM applications as the scene content may change over time due to the presence of dynamic objects, even though the locations are revisited with the same viewpoint. This work enhances the loop closure detection capability of long-term visual SLAM by reducing the number of false matches through the use of location semantics. We extend a semantic visual SLAM framework to build compact global semantic-geometric location descriptors and local semantic vocabulary trees, by leveraging on the already available features and semantics. The local semantic vocabulary trees support incremental vocabulary learning, which is well-suited for long-term SLAM scenarios where the scenes encountered are not known beforehand. A novel hierarchical place recognition method that leverages the global and local location semantics is proposed to enable fast and accurate loop closure detection. The proposed method outperforms recent state-of-the-art methods (i.e., FABMAP2, SeqSLAM, iBOW-LCD, and HTMap) on all datasets considered (i.e., KITTI, Synthia, and CBD), with highest loop closure detection accuracy and lowest query time.