Abstract: Three-dimensional semantic information is an important factor for intelligent machines to understand the world and an important part of artificial intelligence. This paper proposes an improved SLAM framework based on ORB-SLAM2, which can better adapt to the processing of low texture and perceptual aliasing problems in dynamic and complex environments. Combined with the semantic information provided by the convolutional neural network for semantic segmentation, the Bayesian method was used for semantic association to achieve optimized positioning and update in Octomap, and a consistent three-dimensional semantic map was built. The test results based on the public data set show that in a complex environment, the overall mapping accuracy and speed of this method are improved compared with traditional visual SLAM algorithms, and the impact of light transformation is reduced, which has high application value.