TDEG-SLAM: A SLAM System Based on Lightweight Modification of Object Detection and Epipolar Geometry Constraints

Simultaneous Localization and Mapping (SLAM) is a fundamental capability required for intelligent robots to estimate their states in various operating environments. Currently, most visual SLAM systems are designed based on the assumption that the scene is static. Therefore, in scenes with dynamic objects, the system has low robustness and accuracy in state estimation. To solve this problem, we propose a Target Detect Epipolar Geometry SLAM (TDEG-SLAM) system that targets scenes with dynamic objects and combines target detection with epipolar geometry constraints in visual SLAM. TDEG-SLAM is based on the ORB-SLAM2 framework and integrates a lightweight improved YOLOv5 target detection network with a visual SLAM system that uses epipolar geometry constraints for dynamic object detection. Finally, TDEG-SLAM was evaluated on the public dataset from Technische Universität München (TUM), and the results showed that TDEG-SLAM had significantly improved localization accuracy compared to ORB-SLAM2 in scenes with multiple dynamic objects.