Visual SLAM method in dynamic scene

The invention discloses a visual SLAM (Simultaneous Localization and Mapping) method in a dynamic scene. The method comprises the following steps: (1) performing instance segmentation on an original RGB (Red, Green and Blue) picture; (2) carrying out feature point tracking by utilizing an LK optical flow method of the improved image pyramid, and judging parallel dynamic feature points; (3) calculating a basic matrix of two adjacent frames of images, and further screening undetected dynamic feature points by adopting an improved multi-view geometric method fusing dynamic probability; (4) synthesizing a picture without a dynamic object; (5) triangularizing the estimated camera pose to obtain three-dimensional map points, and optimizing the camera pose through a visual re-projection error; (6) optimizing camera poses and map points; (7) loopback detection and loopback optimization are carried out; according to the invention, the precision and robustness of the visual SLAM system in a dynamic scene are improved.