Integrating gray area feature points and static probability to enhance the performance of indoor dynamic visual simultaneous localization and mapping

Abstract In dynamic environments, moving objects pose a great challenge to the accuracy and robustness of visual simultaneous localization and mapping (VSLAM) systems. Traditional dynamic VSLAM methods rely on hand-designed feature frames, and these methods usually make it difficult to fully utilize feature information in dynamic regions. To this end, this paper proposes a SLAM system (GAF-SLAM) that combines gray area feature points, weighted static probabilities, and spatio-temporal constraints. This method realizes the efficient fusion of key point detection and target detection by introducing YOLO-Point to extract gray area feature points from dynamic regions. These feature points are located within the detection frame and have potentially static feature point properties. By combining the reprojection error and polar geometry constraints, potential static feature points are effectively screened out and the identification of these gray area feature points is further optimized. Subsequently, a novel static probabilistic computational framework is designed to assign weights to these gray area feature points and dynamically adjust their influence on the optimization results during the attitude estimation process. By combining static probability with temporal continuity and spatial smoothness constraints, the system achieves significantly improved localization accuracy and robustness in dynamic environments. Finally, the proposed method was evaluated on the TUM RGB-D dataset. The experimental results demonstrate that GAF-SLAM significantly improves pose estimation accuracy and exhibits strong robustness and stability in dynamic indoor environments.