Indoor Topological Mapping Framework for Semantic-enabled Navigation based on Omnidirectional Images

Compared to metric mapping, which often requires detailed information, topological mapping based on human knowledge of the environment is considered a practical method for helping robots recognize spatial environment, thereby providing navigation assistance. This study proposes a novel feature topological mapping technique for indoor navigation, which generates a description of navigation tasks and can be applied to semantic navigation. The framework of the proposed approach comprises two steps: node perception and topological map building. Node perception is performed using a visual landmark detector, which is responsible for identifying specific landmarks in visual data via YOLOv3, and a turning detector, which is responsible for detecting the moving state of the mobile platform using optical flow. Based on the results of node perception, a feature topology map is generated and updated using a hybrid vector of locally aggregated descriptors (VLAD) for loop-closure detection. Further, experiments using a sequence of omnidirectional images are conducted and the results show that the proposed approach can construct the topology of the experimental indoor environment and perform semantic navigation.