Aiming at the problem that the map constructed by traditional visual SLAM does not have semantic information, an octree map construction method based on deep learning algorithm was proposed. In this method, YOLOv7-mask was used to segment the image, the category information of the pixels was obtained, and a semantic grayscale map was generated. Combined with ORB-SLAM2, the new key frame was projected to generate a semantic point cloud map, and finally the semantic point cloud map was converted into a semantic octree map for storage. By building an experimental platform, the semantic octree maps of laboratory scenes under different conditions were constructed, and the feasibility of the proposed method was verified.