Aiming at the problem that visual SLAM is easily affected by moving objects such as vehicles and pedestrians in real environment, resulting in low pose estimation accuracy, a dynamic visual SLAM algorithm combined with lightweight YOLOv5s is proposed. The improved lightweight YOLOv5s is used as an object detection algorithm to judge moving objects. Combined with the proposed method of dynamic feature points elimination, dynamic feature points are eliminated, and only static feature points are used for pose estimation and map tracking. Experiments on TUM data set show that compared with ORB-SLAM3 algorithm, the pose estimation accuracy of the improved algorithm on high dynamic sequence is improved by 89.29%, 65.34% and 94.42% respectively. The results show that the improved algorithm can effectively eliminate dynamic feature points. The pose estimation and positioning accuracy of visual SLAM algorithm in dynamic environment are improved.