A Closed-Loop Detection Algorithm for Indoor Simultaneous Localization and Mapping Based on You Only Look Once v3

Abstract

This paper designs a deep learning-based closed-loop detection algorithm for indoor space. You only look once (YOLO) v3 was adopted to detect the objects in the scene, extract the semantic and position information of the non-dynamic objects contained in the current frame, and solve the similarities between the current frame and key historical frame, thereby completing closed-loop detection. In our network structure, the prior static semantic library is employed to differentiate and eliminate the dynamic objects in the scene, such that the network can apply to most indoor scenes. In addition, the closed-loop detection was made immune to the disturbance of dynamic objects. The extracted semantic information can be applied to modules like visual odometer and semantic maps.