This paper presents a review of approaches to the automatic detection of anomalous human behavior from surveillance video recordings. The relevance of the study is обусловлена the diversity of existing yet insufficiently systematized approaches to anomaly detection, as well as the presence of unresolved problems that remain in this field. The review focuses on methods aimed at detecting deviant human behavior. Existing studies are systematized, the main directions of the field’s development and its current challenges are identified, and the advantages and limitations of the considered approaches are analyzed. Special attention is given to the datasets used for human behavior anomaly detection, including their application focus, data volume, and annotation characteristics. The study shows that semi-supervised learning currently dominates in the VAD field, whereas supervised learning remains relevant for narrow domains in which anomalous behavior can be clearly defined.