Stable Observation Generation of Retroreflective Landmarks for Utility-Tunnel Inspection Robots

Utility tunnels are narrow, repetitive, and GNSS-denied, while reflective infrastructure can create false landmark detections. This paper presents a lightweight observation method for low-cost retroreflective landmarks on inspection robots. The first contribution combines intensity filtering with local geometric verification to suppress reflective-background interference. The second combines sparse-candidate recovery with temporal tracking and state confirmation to retain stable observations under sparse returns and short occlusions. On annotated real LiDAR data, the method achieves 88.7% precision, an F1 score of 80.3%, and a mean processing time of 1.57 ms. The resulting observations are associated with a pre-built reflector map; their real-site contribution to absolute localization accuracy requires external trajectory ground truth and is not claimed here.