In order to improve the accuracy and real-time of the loop detection module of mobile robot, an improved loop detection algorithm based on feature map is proposed. On the basis of traditional model, selecting the feature points of key frames and efficient feature points to build feature maps, using visual dictionary tree to description scene of feature map and key frame. Secondly, to improve the bag of words model, scene segmentation is applied to image information extraction and feature clustering. Finally, a visual dictionary tree based on hierarchical K++ means is established, and an improved method of score matching based on hierarchical pyramid TF-IDF (term frequency-inverse document frequency) is obtained. The test results show that compared with FAB-MAP (fast appearance-based mapping) and RGB-D SLAM v2, the improved algorithm has better performance in feature point size, real-time performance and recall rate.