Abstract: To address the limitations in semantic context utilization and spatio-temporal feature modeling in existing anomaly detection methods, a video anomaly detection method based on video caption augmentation and dual-stream feature fusion was proposed. Video captions were automatically extracted and encoded using the contrastive language-image pre-training(CLIP)model to serve as auxiliary semantic context information for anomaly detection. A spatio-temporal adaptive embedding module was introduced to capture subtle temporal variations and complex spatial structures within videos, enabling effective spatio-temporal feature fusion. A cross-modal alignment module was further designed to deeply integrate contextual semantic features with spatio-temporal visual features, allowing more accurate capture of joint spatio-temporal-semantic representations of anomalous events. Experimental results showed that the method achieved area under the curve A_UC scores of 97.54% on the ShanghaiTech dataset and 90.54% on the CUHK Avenue dataset. The results confirmed the performance and robustness of the method across multiple public video anomaly detection datasets, providing an effective solution for this critical task.

Key words: video anomaly detection, video caption, spatio-temporal adaptive embedding, temporal Transformer, spatial Transformer