Monocular visual SLAM using MR-EKF in repetitive pattern environment

Abstract

In order to automate industrial inspection robots in repetitive pattern environments, robot localization and map building are necessary. A viable solution to address this issue is Simultaneous Localization and Mapping (SLAM), with visual SLAM specifically utilizing camera images. However, conventional visual SLAM techniques struggle in tile environments due to errors in feature matching caused by the presence of recurring similar patterns. To tackle this issue, a visual SLAM system based on Map Point Relation Extended Kalman Filter is proposed. The proposed system employs vertices of the patterns as map points for visual SLAM. The patterns in the image are detected using a Convolutional neural network based object detector. Subsequently, the vertices of the patterns are computed through postprocessing. The system tracks the detected patterns and matches their vertices for each image frame, effectively addressing feature matching errors that emerge in environments with repetitive patterns. The proposed Map Point Relation Extended Filter is an adaptation of the Extended Kalman Filter, with the inclusion of error covariance for map point relations, rendering it specifically tailored for repetitive pattern environments. Map Point Relation refers to the positional relationship between map points, established in this study based on the characteristics of repetitive pattern environments. Performance experiments conducted in simulation-based repetitive pattern environments demonstrate the effectiveness and noise robustness of the proposed system. Furthermore, the applicability of the proposed system to real-world robotic systems is verified through a simulation experiments with a shape similar to the actual industrial environment.