HIERARCHIAL TRANSFORMER NETWORKS FOR ANOMALY DETECTION IN SURVEILLANCE VIDEOS

Gultiaev Andrey A.

This paper presents a novel approach for anomaly detection in surveillance videos using hierarchical transformer networks without relying on convolutional neural networks. We leverage Video Vision Transformers (ViViT) combined with contrastive learning to extract meaningful embeddings from video segments. To handle variable-length video clips, we introduce a hierarchical transformer architecture that captures both segment-level and event-level representations. Trained on the DCSASS dataset, our method demonstrates significant improvements in classification, clustering, and anomaly detection tasks compared to traditional approaches. Our results indicate that the proposed model can effectively assist surveillance operators in detecting abnormal activities, thereby enhancing security measures.

Keywords:machine learning, artificial intelligence, computer vision, neural network, transformer, contrastive learning, vector embedding, classification, clustering, anomaly detection