Abstract: An improved loop closure detection algorithm based on historical model set is proposed to improve the precision ratio and real-time performance of loop closure detection in mobile robot SLAM(simultaneous localization and mapping). First of all, the descriptor of feature is added to the model on the basis of frame-to-model registration model based on Kinect sensor, and it is updated with a weighting method to construct historical model set. A visual vocabulary tree is used for scene description of the historical model set and current frame. Then the pyramid TF-IDF(term frequency-inverse document frequency) scoring match scheme is improved by using inverse proportional function instead of minimum function to optimize the similarity score at a single node between two images. On one hand, the perceptual aliasing can be reduced with the improved scoring match scheme, and thus the precision ratio of loop closure detection is increased. On the other hand, by using the improved scoring match scheme, the affiliation between current frame and historical model set can be judged effectively with fewer number of comparison than the traditional frame-by-frame comparison method, so that the real-time performance of loop closure detection is improved. Next, the improved pyramid TF-IDF scoring match scheme is utilized for the similarity analysis between current frame and the keyframes affiliated to the candidate historical model set to extract candidate loop closure. Finally, the outliers in the candidate loop closure are discarded according to temporal continuity and epipolar geometry constraints. Contrast experiments based on the datasets and the actual scene show that, compared with the IAB-MAP(incremental appearance-based mapping), FAB-MAP(fast appearance-based mapping) and RTAB-MAP(real-time appearance-based mapping), the improved loop closure detection algorithm proposed has better real-time performance and higher recall ratio in the case of 100% precision ratio.