Much of the existing SLAM system in dynamic scenes has tended to focus more on improving the accuracy of localization at the expense of real-time performance. This paper proposes a real-time accurate dynamic SLAM system based on ORB-SLAM3. It combines the latest efficient instance segmentation network YOLOv8s-Seg and geometric constraints to extract dynamic objects and filter dynamic feature points respectively. To balance accuracy and efficiency, a new efficient strategy of performing instance segmentation on keyframes while object tracking and prediction on others is proposed in parallelized semantic threads, which greatly reduces the number of forward inferences of the network, saves computational overhead, and shortens the time delay. In addition, to avoid the disadvantages of difficulty in filtering the features due to the small motion amplitude between adjacent frames, this paper introduces an inter-frame epipolar line constraint checking mechanism based on sliding window to judge the motion state of each feature in object regions as accurately as possible, to retain more static features and improve the accuracy of pose estimation. Experimental evaluation is conducted on dynamic sequences of the TUM RGB-D dataset. The results show that the proposed method enables the system to reduce the absolute trajectory error (ATE) and relative attitude error (RPE) significantly compared to the baseline, and it also ranks high compared to several existing advanced algorithms. Further, it performs well in real-time, which fully demonstrates that the proposed system excels in both accuracy and efficiency.