Most traditional visual simultaneous localization and mapping (SLAM) systems typically assume a static environment; however, real-world environments often contain moving objects and obstacles, leading to a significant number of mismatched and dynamic points which can degrade localization accuracy. This paper proposes a semantic vSLAM system based on the ORB-SLAM3 framework and deep learning techniques, integrating object detection and optical flow methods to improve localization accuracy in dynamic environments. Firstly, an enhanced YOLOX-S object detection algorithm is utilized to identify potential dynamic targets. Subsequently, a combination of geometric and optical flow methods is employed to precisely detect outliers, with continuous adjustments to dynamic bounding box thresholds based on the motion states of objects and humans. Ultimately, points within static bounding boxes retained in dynamic frames are preserved, while others within dynamic frames are eliminated. The system′s accuracy is evaluated using the TUM and KITTI datasets. Experimental results demonstrate that under highly dynamic sequences, the proposed system achieves an average reduction of 69.26% and 16% in root mean square error of absolute trajectories compared to ORB-SLAM3 and Crowd-SLAM, respectively, and a 15% average improvement in localization accuracy in dynamic scenes when compared to DynaSLAM, thereby validating the enhanced system performance in dynamic environments.Moreover, the results of real-world scene tests demonstrate that the system performs well in various complex environments.