The Simultaneous Localization And Mapping (SLAM) is a difficult problem in the field of robotics. Rao-Blackwellized particle filters algorithm is widely used to solve this problem. In the traditional implementation, the proposed distribution with high error will calculate a large number of sampled particles to fit the target distribution. Frequent resampling steps will lead to gradual dissipation of particles and waste a lot of computing resources. In this study, the motion model and observation information are combined to optimize the proposed distribution, reduce the number of sampled particles, and the adaptive resampling method is introduced to reduce the steps of resampling. In the implementation of the algorithm, the tree data structure is used to store the environment map. The experimental results show that the improved algorithm can significantly improve the computational efficiency, reduce the storage consumption, and build more accurate map.