DVSG-SLAM: Dynamic Visual SLAM Based on semantic and geometric information

To address the challenge that dynamic visual SLAM is easily affected by dynamic objects in real environments, resulting in reduced pose estimation. This paper introduces a dynamic visual SLAM algorithm DVSG-SLAM that combines semantic information and geometric information. Based on ORB_SLAM3, the semantic information part is integrated with the target detection network YOLOv5s and the Lucas-Kanade sparse optical flow method. The geometric information part is integrated with LSD line features and ORB point features, where point features provide dense geometric information and line features supplement the extraction of stable structured geometric constraints. This paper uses the TUM dataset developed by the Technical University of Munich for experiments. In a dynamic environment, the root mean square error of this algorithm is improved by an average of 76.55% compared with ORB_SLAM3. Experiments show that this algorithm can effectively remove dynamic obstacles and enhance the pose estimation accuracy of the SLAM system.