Method of the automated human anomalies behaviour detection without prior training

This paper presents a method for automatic detection of anomalous human movements in video recordings without prior training. The proposed approach includes input data preprocessing, motion segmentation using the PELT algorithm, and clustering of the segments with the DBSCAN method. The input dataset is prepared using neural networkbased detectors. Automatic parameter selection for clustering is performed using the elbow method, enabling the separation of typical and anomalous motion patterns. Experimental results on the MOT-17 dataset demonstrate the capability of the method to identify anomalous movements even in the presence of noise and missing data. Evaluation metrics, including the Calinski – Harabasz index, the Davies – Bouldin index, and the silhouette coefficient, indicate an acceptable clustering quality, which confirms the potential of the proposed method for application in video surveillance systems.