Simultaneous Localization and Mapping (SLAM) remains a key focus in robotics research, significantly contributing to positioning and navigation in uncharted environments. Although many existing SLAM algorithms excel in static conditions, their performance often falters in dynamic settings. While the incorporation of deep learning methods has greatly enhanced accuracy, achieving real-time performance continues to be a challenge, particularly for embedded devices with limited resources. This study introduces a dynamic SLAM framework leveraging an RGB-D camera, integrating the FastestDet object detection model with ORB-SLAM3 to enable effective segmentation and tracking of dynamic and static regions. The system introduces a dedicated object detection thread, coupled with an adaptive depth estimation algorithm, to segment dynamic objects and subsequently remove dynamic points. An adaptive mask compensation tracking module enables the tracking thread to operate independently of object detection results, enhancing efficiency. Optical flow is employed for non-keyframe tracking, substantially improving real-time performance. To further enhance system accuracy, a method integrating inverse optical flow with epipolar constraints and motion consistency filtering, alongside a dynamic point filtering module based on clustering and optical flow weighting, is proposed. The presented method was tested on the TUM and Bonn datasets. In comparison with ORB-SLAM3, it achieves a 97.82% reduction in average absolute trajectory error in highly dynamic environments, enhances frame processing speed by 72.99%, and attains a tracking frame rate of 157 FPS on a typical laptop.