Research on Visual SLAM Method Based on Deep Learning in Dynamic Environments

Simultaneous Localization and Mapping (SLAM) refers to the technology that enables a mobile robot operating in an unknown environment to utilize its onboard sensors to accomplish real-time mapping of the environment while simultaneously achieving autonomous localization. Compared to laser-based SLAM, visual SLAM has become a research focus due to the advantages of camera sensors, such as low cost, compact size, light weight, and rich information content. However, the current mainstream visual SLAM system development is based on the assumption of static environments, neglecting the presence of dynamic objects. To address the issues of poor robustness and localization accuracy of traditional feature-based SLAM systems in dynamic environments, a dynamic SLAM method for dynamic indoor scenes is proposed. This paper improves upon the ORB-SLAM2 algorithm framework by using the YOLOV8s-seg instance segmentation model to detect dynamic objects in actual scenes and discriminates feature points on dynamic objects, and employs the deep learning descriptor BEBLID to replace the original ORB descriptor to resolve the problem of low feature matching rates. Finally, experiments conducted on the TUM dataset demonstrate that, compared to the original ORB-SLAM2 algorithm, our algorithm effectively improves the positioning accuracy and robustness of the visual SLAM system on highly dynamic datasets, with an absolute RMSE improvement of 96.25% in trajectory error.