Topological Mapping with Constrained Optimization based on Visual Place Recognition and Orientation Constraints

This paper proposes a method for topological mapping that guarantees orientation accuracy. Unlike pure topological maps, the orientations of arcs connecting nodes in our topological maps match those in real environments. The proposed method uses the Visual Place Recognition (VPR) method, AnyLoc-VLAD-DINOv2, to extract global descriptors from images obtained by a 360-degree camera, and then creates nodes and detects loops based on these descriptors. Furthermore, the initial orientation of each node is calculated using angular velocity obtained by an IMU. The proposed method creates two types of orientation constraints. Heading constraints are created from initial orientations, and loop constraints are created from loop detection. Subsequently, node orientations and arc lengths are corrected through constrained optimization with two types of orientation constraints. Through indoor and outdoor experiments, the proposed method enabled topological mapping with both topological consistency and ori- entation accuracy. Nodes of the topological maps were created adaptively based on the appearance of each location within the environments, and the node spacing varied accordingly. Furthermore, through constrained optimization with two types of orientation constraints, node loops were closed and arc orientations matched those in real environments. Even in environments containing multiple loops, the proposed method enabled topological mapping while simultaneously satisfying the constraint of each loop.