Abstract: Firstly, the existing scene appearance description methods and their performances when dealing with various unstructured factors are discussed. Secondly, the characteristics of scene memory models based on pure image retrieval, and topological information and metric information are introduced, and as a case, the performance optimization and construction methods of the visual dictionary are discussed. Thirdly, three key steps of loop closure detection (LCD) are analyzed, including loop closure matching, subsequent verification, and optimization. Fourthly, the popular evaluation metrics and the standard datasets for LCD methods are presented. Finally, the current situation of LCD research is summarized, and the potential of novel techniques including unsupervised learning, semantic context, and lightweight model is described.