Reflective Surface Detection and Its Impact on SLAM Trajectory Accuracy

Abstract

Simultaneous Localization and Mapping (SLAM) algorithms, such as ORB-SLAM3, have advanced significantly in recent years, enabling more accurate and efficient navigation for mobile robots in diverse environments. However, reflective surfaces, particularly mirrors, continue to pose critical challenges, causing localization errors and distorted maps due to their disruptive effects on point clouds and visual data. This study introduces a novel method leveraging machine learning for mirror detection using a camera, which also serves as the primary sensor for localization, thereby reducing system complexity and cost. Detected reflective surfaces are corrected in real time by dynamically removing the corresponding distorted regions from the map, allowing ORB-SLAM3 to continue mapping accurately. The proposed approach achieves approximately 97% accuracy in mirror detection and effectively addresses the disruptive effects of both framed and frameless mirrors. Experimental results in diverse indoor scenarios demonstrate the potential of this method to improve mapping accuracy and reliability, providing a scalable solution for real-time autonomous navigation in environments with reflective surfaces.