Abstract: The traditional simultaneous localization and mapping (SLAM) algorithm suffers from excessive rigidity and redundancy in keyframe selection， and poor real-time performance and robustness in complex dynamic SLAM scenarios. We propose a lightweight, semi-direct, dynamic RGB-D SLAM algorithm to address these issues. First, we introduce the multi-level FAST (features from the accelerated segment test) feature removal algorithm, which combines a sparse alignment method with the Shi-Tomasi scoring function. This approach eliminates dynamic object feature points, reduces computation time, and improves robustness in dynamic environments. Next, we propose an adaptive keyframe selection threshold strategy, which dynamically adjusts the threshold based on inter-frame information difference, enhancing positioning accuracy while reducing redundant keyframes. Finally, we incorporate a dense mapping module capable of removing dynamic objects while preservingthe static environment in dynamic scenes. Experimental results on the TUM dynamic dataset and OpenLORIS-Scene dataset demonstrate that, in low-dynamic environments, the proposed algorithm improves accuracy by approximately 47% compared to ORB-SLAM3 and by approximately 90% in high-dynamic environments. Compared with four dynamic SLAM algorithms, our method reduces absolute trajectory error, mapping error, and average time per frame by at least 23%, 42% and 26%, respectively. The dense mapping module also fully reconstructs the real scene with improved accuracy.