Abstract: A dynamic obstacle avoidance method based on real-time local grid map construction includes: acquiring and inputting Red-Green-Blue-RGBD image data of a real indoor scene into a trained obstacle detection and semantic segmentation network to extract obstacles of different types and semantic segmentation results in the real indoor scene and generate 3D point cloud data with semantic information; according to the 3D point cloud data, extracting and inputting state information of a dynamic obstacle to a trained dynamic obstacle trajectory prediction model, and predicting a dynamic obstacle trajectory in the real indoor scene to build a local grid map; and based on a dynamic obstacle avoidance model, sending a speed instruction in real time to the mobile robot to avoid various obstacles during the navigation process.