Localization is a foundational task for mobile robot navigation in an indoor environment. In order to accommodate more ordinary indoor environment, improved adaptive particle filters is proposed in this paper, which can adjust the number of samples adaptively according to the status of sample convergence. We distribute random samples based on the analysis result of the sensor reading in the initial state and choose the observed features from geometric feature-based prior map instead of the bins in the grid map. Experiments result implemented in real mobile robot platform and further localization result analysis show the validity and practicability of this improved method