Abstract: In view of the fact that the Gmapping SLAM algorithm has high requirements on the accuracy of odometry positioning information in the process of map construction， and there are problems such as particle dissipation and degradation， Firstly， a parallel visual recognition and localization network is designed to strengthen the localization ability and improve the semantic information and composition accuracy； Secondly， the optimization proposal distribution is improved， we use the laser observation model to replace the odometrg motion model and perform Gaussian sampling to cover the probability distribution of the robot with fewer particles； Finally， the visual information and laser information are fused by Bayesian rule， and the original algorithm is compared. The results show that the algorithm improves the accuracy and robustness of map construction and enriches the semantic information.

Key words: SLAM, computer vision, laser observation model, neural networks, bayesian fusion