Simultaneous localization and mapping (SLAM) serves as a fundamental capability for autonomous systems to explore unfamiliar surroundings by estimating their position through sensor data. Among these SLAM systems, visual SLAM (vSLAM) has been widely explored due to the cost-effectiveness and rich data of cameras, and it excels at building accurate maps in static environments. However, moving objects in dynamic environments may occlude static scenes, hindering feature extraction, disrupting pose estimation, and causing accumulated errors that ultimately risk system failure. To address these issues, we propose a real-time RGB-D vSLAM system that can effectively mitigate dynamic object interference. This proposed system integrates coarse masks along with rigid body constraints to remove dynamic features. Moreover, an improved bag-of-words mechanism that leverages multi-frame positional data and visual features is developed to enhance loop detection accuracy and optimize map construction. Finally, this system employs a multi-threaded architecture to boost the system’s processing performance. Experimental results demonstrate that the proposed system achieves significant improvements in accuracy while maintaining computational efficiency.