<p>Addressing the challenge of visual simultaneous localization and mapping (SLAM) in dynamic environments remains a significant research focus. Current methods often suffer from inaccurate initial position estimations and the presence of unknown objects. To tackle these issues, we propose a real-time RGB-D SLAM system based on local observation consistency, which circumvents the limitations of feedback-based approaches. To reduce computational burden, we implement a grid-based K-means clustering method for spatial partitioning. Additionally, a two-stage probabilistic quantification model utilizing the median absolute deviation is proposed to evaluate the dynamic extent of each region, thereby enhancing the system robustness in detection. Given the continuity of motion, we develop a probabilistic grid iterative model with historical information storage capabilities, significantly improving the ability of system to rapidly and accurately identify dynamic areas. We validated the effectiveness of our method through a series of experiments on the TUM RGB-D benchmark dataset and real-world tests on low-power embedded devices. The experiment video: <a href="https://youtu.be/CblTuw65Mlc">https://youtu.be/CblTuw65Mlc</a></p>