Visual SLAM Based on Geometric Constraints and Lightweight Object Detection Network in Dynamic Environments

In dynamic environments, performing Simultaneous Localization and Mapping (SLAM) is often disrupted by moving objects, resulting in decreased localization accuracy and robustness. This paper proposes a dynamic feature point elimination algorithm based on geometric constraints and an object detection network. By using only static feature points in the camera pose calculation, the algorithm processes images captured by visual sensors, effectively enhancing the system's localization accuracy and robustness in dynamic scenes. Additionally, to address the issues of large parameter sizes and the difficulty of meeting SLAM's real-time requirements in current object detection networks, this paper presents a lightweight improvement of YOLO-V5s. In the TUM dynamic scene dataset, the improved algorithm achieves a 95.29% increase in localization accuracy compared to ORB-SLAM2 and processes 13.5 frames per second, meeting the real-time requirements of visual SLAM.