Video anomaly detection via cross-modal fusion and hyperbolic graph attention mechanism
To address the challenges of modality information imbalance, non-uniform audiovisual noise, and modality asynchrony in video anomaly detection, a multimodal video anomaly detection method called CM-HVAD was proposed for accurate anomaly detection. Firstly, a novel dynamic cross-modal fusion module w...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | Chinese |
| Published: |
Editorial Department of Journal on Communications
2025-06-01
|
| Series: | Tongxin xuebao |
| Subjects: | |
| Online Access: | http://www.joconline.com.cn/zh/article/doi/10.11959/j.issn.1000-436x.2025110/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | To address the challenges of modality information imbalance, non-uniform audiovisual noise, and modality asynchrony in video anomaly detection, a multimodal video anomaly detection method called CM-HVAD was proposed for accurate anomaly detection. Firstly, a novel dynamic cross-modal fusion module was introduced to dynamically compress and reweight multimodal features through autonomous learning of cross-modal weights, thereby achieving balanced and enhanced fusion of visual and audio features. Secondly, to address the issue of modal asynchrony in multimodal data, a modal consistency alignment module was proposed, which aligned modal semantics along the temporal frame sequence to ensure both temporal and semantic consistency in multimodal data. Finally, a hyperbolic graph attention mechanism was incorporated to effectively capture the hierarchical relationships between normal and abnormal representations through the pattern separation property of hyperbolic space, thereby improving detection accuracy. The results show that the proposed method achieves 86.47% AP on XD-Violence and 87.12% AUC on UCF-Crime, outperforming baseline methods. |
|---|---|
| ISSN: | 1000-436X |