Visual Simultaneous Localization and Mapping (VSLAM) technology is widely applied, but traditional algorithms typically assume a static environment and perform poorly in complex dynamic scenarios, especially when multiple objects exhibit local or intermittent motion. To address this issue, we propose the MAS-VINS algorithm, which integrates Multi-Scale Local Outlier Factor (MSLOF) and dynamic factor graph optimization. This algorithm combines an improved object detection network with depth information to perform lightweight segmentation, incorporating semantic information while balancing real-time performance and accuracy. Additionally, the MSLOF algorithm, combined with the moving consistency check mechanism, identifies dynamic features in real-time and discerns the motion patterns of objects. Finally, the algorithm dynamically adjusts the weight network in factor graph optimization to achieve efficient pose estimation and optimization. Experimental results show that MAS-VINS achieves superior localization accuracy and adaptability across various dynamic environments.