Semantic Visual SLAM Based on Target Detection Algorithm in Indoor Dynamic Scenes

To solve the problem of decreased positioning accuracy and poor robustness of simultaneous localization and mapping (SLAM) systems based on static environment assumptions when processing dynamic objects, a semantic vision SLAM system for indoor dynamic environments was proposed, which was based on ORB_SLAM2, and added with a dynamic target detection thread. While extracting ORB features from the input image, YOLOv5s network was used for dynamic target detection, combining optical flow method and epipolar geometric constraints to screen dynamic feature points, and finally static feature points were used for pose estimation. Using the data set of Technical University of Munich (TUM) to compare with ORB_SLAM2, results show that the system significantly reduces the trajectory error. Compared with the system in the dynamic environment such as DS-SLAM and DynaSLAM, this system can effectively balance the accuracy, robustness and rapidity of the pose estimation of the semantic visual SLAM system.