期刊: [Institution of Engineering and Technology] 日期:2025-08-01卷期号:2025 (15): 418-420
标识
DOI:10.1049/icp.2025.2583
摘要
We propose a novel weakly supervised framework for video anomaly detection that integrates multi-modal features from the Inflated 3D Convolutional Network (I3D) and Contrastive Language-Image Pre-training (CLIP) via a multi-head cross-attention mechanism. The framework further incorporates dynamic top-K selection, temporal context aggregation, and comprehensive visual prompt learning—including Anomaly-Focused, Normal, Episodic, and Attention Aggregation (AA)-Semantics Prompts, along with a Textual Prompt. These components enhance temporal coherence, semantic alignment, and anomaly discrimination, enabling more accurate detection in long video sequences. Experiments on UCF-Crime and XD-Violence demonstrate competitive performance, validating the effectiveness of the proposed approach.