When the SLAM system estimates the camera position, a large number of feature points of moving objects participate in the feature tracking thread leading to a decrease in the accuracy and robustness of the algorithm, so how to efficiently and accurately reject the dynamic objects in the scene is particularly important. Existing dynamic vision SLAM algorithms may miss detecting or incorrectly recognize static objects as dynamic objects and reject them when dealing with dynamic objects, which triggers the problem of insufficient number of static feature points, thus affecting the stability and accuracy of the SLAM system. Therefore, this paper proposes a visual SLAM method based on panoptic segmentation and multi-view geometry, which uses panoptic segmentation FPN network to accurately recognize all objects in the segmented image, rejects a priori dynamic feature points and retains as many static features as possible, based on which LK optical flow method with fused image pyramid is used to realize optical flow tracking and reject parallel dynamic feature points, and potential dynamic feature points are used to track the dynamic feature points. The potential dynamic feature points are rejected more effectively by the multi-view geometry method based on dynamic probability, which avoids the omission of dynamic feature points and realizes the comprehensive screening of dynamic objects in the scene to improve the accuracy of the system. The construction of semantic map and octree map is realized on the basis of sparse point cloud constructed by the system. The experiments use the TUM RGB-D dataset to verify the system localization accuracy, and the results show that the root mean square error (RMSE) of the absolute trajectory error of this algorithm is reduced by an average of 84.34% in all sequences compared with ORB-SLAM2, which significantly improves the robustness and accuracy of the system,and it is of use to construct two maps that can be used for SLAM upper layer tasks.