Abstract: The performance of visual environment modeling in appearance-based robot loop closure detection by using conventional vocabulary is restricted by limited number of visual words and high computational cost. We construct a visual vocabulary tree by clustering the visual features hierarchically captured by a mobile robot. The TF-IDF entropy for each node is computed and is treated as the weight of each visual word, and the inverted index of image-word is exploited. To avoid the quantization error of single scale vocabulary and the neglect of the different discriminative power among different level words of tree-based match, we take advantage of the robustness of high level words and the discriminability of low level words to present a pyramid scoring match scheme. The candidates of loop closures are detected by using a similarity threshold. A posteriori management helps discard outliers by verifying that the two images of the loop closure satisfy some hypothesis constraints. The experiments of loop closure detection in mobile robotics demonstrate that our scheme improves similarity calculation significantly in both accuracy and efficiency and obtains a higher precision-recall ratio with a faster speed of loop closure detection compared to the traditional methods.