Panoptic-Guided Scene Description for Visual Place Recognition of Patrol Robotic Systems

This paper addresses the challenge of visual place recognition (VPR) for patrol robots, aiming to estimate coarse location in GPS-denied environments. Place recognition is essential for autonomous navigation, particularly to avoid kidnapped robot scenarios. Since GPS signals are often unstable in cluttered environments such as urban areas and industrial complexes, many patrol robots rely on various sensors. Among them, VPR (i.e., image-based place recognition) is especially practical due to the versatility of camera sensors, which are lightweight, low-cost, and power-efficient. However, VPR performance can degrade significantly under large viewpoint variations between query and reference images. To address this, we propose a novel panoptic-guided VPR method that enhances robustness to such changes. The method first employs panoptic segmentation to extract pixel-level class information, then constructs a class frequency histogram while discarding predefined dynamic object classes to improve stability. In addition, a topological graph is generated to model objects and their relative spatial configuration, helping mitigate viewpoint-induced mismatches. The similarity between the query and reference candidates is computed using a radial basis function (RBF) for the histogram comparison and the Chebyshev distance for graph embedding comparison. Experimental results in multiple benchmark data sets demonstrate that our panoptic-based descriptor achieves an approximately 10% improvement in VPR performance over state-of-the-art baselines.