Visual SLAM based on semantic segmentation and optical flow in dynamic scenes

Visual simultaneous localization and mapping (SLAM) systems that assume static environments often struggle with dynamic objects, resulting in degraded localization robustness. To address this limitation, we propose a novel dynamic SLAM system for indoor environments, integrating the YOLOv9 instance segmentation network with the Unimatch optical flow model. Our method effectively identifies and handles moving objects in dynamic scenes, thereby enhancing localization and mapping performance. Specifically, YOLOv9 is employed to generate preliminary masks of potentially dynamic objects, while Unimatch provides dense optical flow information for inter-frame motion analysis. Then, the weights of feature points were defined, and the feature point weights were initialized based on mask information. A joint optimization method based on weights was proposed to optimize the camera pose. Finally, a dense point cloud map of the static environment is generated by filtering out dynamic elements. Experiments conducted on the Bonn and TUM datasets demonstrate that our method significantly outperforms ORB-SLAM2 in highly dynamic scenes, achieving superior localization accuracy and more consistent map reconstruction.