Aiming at the problem that adaptive Monte Carlo localization (AMCL) algorithm is difficult to localize in large scenes and similar environments. This paper uses a semantic information-assisted approach to improve the AMCL algorithm. This method realizes the robust localization of the robot in the large scenes and similar environments. Firstly, the 2D grid map created by simultaneous localization and mapping using lidar can obtain highly accurate indoor environmental contour information. Secondly, the semantic object capture is achieved by using a depth camera combined with an instance segmentation algorithm. Then, the semantic grid map is created by mapping the semantic point cloud through the back-projection process of the pinhole camera. Finally, semantic grid maps are used as a priori information to assist in localization, which will be used to improve the initial particle swarm distribution in the global localization of the AMCL algorithm and thus will solve the robot localization problem in this environment. The experimental evidence shows that the semantic grid map solves the environmental information degradation problem caused by 2D lidar as well as improves the robot’s perception of the environment. In addition, this paper improves the localization robustness of the AMCL algorithm in large scenes and similar environments, resulting in an average localization success rate of about 90% or even higher, and further reduces the number of iterations. The global localization problem of robots in large scenes and similar environments is effectively solved.