Abstract: In order to improve the positioning accuracy and real-time performance of visual simultaneous localization and mapping (vSLAM) algorithm in dynamic scenes, a real-time semantic SLAM algorithm based on optical flow is proposed. In order to improve the positioning accuracy of the algorithm, SegNet image segmentation network and PWC-Net optical flow network are used to remove dynamic feature points. At the same time, the speed of feature points is estimated through optical flow information, which is used as a constraint to remove outliers. Meanwhile, a segmentation strategy is proposed that can obtain more semantic information in a limited amount of time. The experimental results show that compared with other vSLAM algorithms, our algorithm effectively reduces the error of SLAM camera pose estimation in dynamic environments, and improves real-time performance.