Source-linked AI summary
EventVAD: Training-Free Event-Aware Video Anomaly Detection
Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, Xiaochen Wang, Hao Tang, Yan Wang, Shuyan Li
TL;DR
Training-free VAD methods face difficulty modeling temporal dynamics and diverse events while efficiently localizing fine-grained anomalies in long videos. EventVAD segments videos into coherent events, models temporal features with optical-flow-enhanced dynamic graphs, and uses MLLMs to score anomalies within events. EventVAD achieves SOTA on UCF-Crime and XD-Violence in the training-free setting with 7B MLLMs, compared with baseline 13B-or-larger MLLMs.
Problem
Training-free VAD methods face difficulty modeling temporal dynamics and diverse events while efficiently localizing fine-grained anomalies in long videos.
Method
EventVAD segments videos into coherent events, models temporal features with optical-flow-enhanced dynamic graphs, and uses MLLMs to score anomalies within events.
Results
EventVAD achieves SOTA on UCF-Crime and XD-Violence in the training-free setting with 7B MLLMs, compared with baseline 13B-or-larger MLLMs.
Takeaways & Limitations
Event-centric temporal reasoning enables accurate training-free anomaly detection while reducing model size and associated storage and inference costs.
Abstract
from arXiv · showhide
Video Anomaly Detection~(VAD) focuses on identifying anomalies within videos. Supervised methods require an amount of in-domain training data and often struggle to generalize to unseen anomalies. In contrast, training-free methods leverage the intrinsic world knowledge of large language models (LLMs) to detect anomalies but face challenges in localizing fine-grained visual transitions and diverse events. Therefore, we propose EventVAD, an event-aware video anomaly detection framework that combines tailored dynamic graph architectures and multimodal LLMs through temporal-event reasoning. Specifically, EventVAD first employs dynamic spatiotemporal graph modeling with time-decay constraints to capture event-aware video features. Then, it performs adaptive noise filtering and uses signal ratio thresholding to detect event boundaries via unsupervised statistical features. The statistical boundary detection module reduces the complexity of processing long videos for MLLMs and improves their temporal reasoning through event consistency. Finally, it utilizes a hierarchical prompting strategy to guide MLLMs in performing reasoning before determining final decisions. We conducted extensive experiments on the UCF-Crime and XD-Violence datasets. The results demonstrate that EventVAD with a 7B MLLM achieves state-of-the-art (SOTA) in training-free settings, outperforming strong baselines that use 7B or larger MLLMs.
1 INTRODUCTION
EventVAD addresses the difficulty of training-free video anomaly detection in modeling temporal dynamics and localizing fine-grained abnormal events. It segments videos into events, refines temporal features with optical flow and graph attention, and achieves SOTA results with a smaller model.
- Motivation: Training-free VAD avoids task-specific training but existing methods struggle with unseen settings, temporal dynamics, diverse events, and frame-level consistency.Supervised methods require in-domain fine-tuning, while image-based MLLMs and redundant VQA pipelines limit temporal modeling and efficiency.
- Proposed framework: EventVAD segments long videos into short events to improve temporal consistency during MLLM-based anomaly scoring.The framework is designed to reduce scoring errors while localizing fine-grained abnormal frames.
- Proposed framework: Optical-flow-enhanced frame features and Graph Attention aggregate temporal information to detect event boundaries and segment events more accurately.RAFT captures fine-grained inter-frame features, while Graph Attention refines temporal features across frames.
- Results: EventVAD achieves SOTA across UCF-Crime and XD-Violence in training-free settings and outperforms some weakly supervised methods.Its 7B-parameter model is smaller than 13B-parameter baselines, reducing storage and inference costs.
2 RELATED WORK
Prior VAD research includes weakly supervised and unsupervised training paradigms, while training-free approaches use language models to avoid retraining. EventVAD extends this direction through temporal segmentation into discrete event units for long-form video analysis.
- 2 RELATED WORK: Weakly supervised and unsupervised VAD reduce annotation requirements but training-based approaches still require data collection and model retraining for deployment.Weakly supervised methods use video-level labels, whereas unsupervised methods model normal video structure without labeled data.
- 2 RELATED WORK: EventVAD uses temporal segmentation into discrete event units to mitigate error propagation and reduce false detections in long-form video analysis.This event-based strategy builds on language-model approaches to training-free VAD.
- 2 RELATED WORK: Video-LLM systems commonly connect visual features to language models through adapters that compress frame-level representations into fixed-length tokens.Examples include cross-attention, Q-Former, linear projection, and dynamic visual tokenization.
3 METHODOLOGY
EventVAD models videos with dynamic multimodal spatiotemporal graphs, propagates features with orthogonal attention, detects event boundaries statistically, and scores anomalies within semantic event units using MLLMs.
- 3.1 Event-Aware Dynamic Graph Construction: Semantic and motion features are fused into frame nodes, while dynamically established edges encode cross-modal similarity and time-sensitive spatiotemporal relationships.The semantic branch uses normalized CLIP embeddings, and the motion branch uses RAFT optical flow; temporal decay suppresses remote associations.
- 3.2 Graph Attention Networks Propagation: Orthogonally constrained graph attention propagates information across temporally connected neighbors while preserving global divergence and local consistency.Orthogonal projection prevents dimension collapse, and temporal connectivity limits attention to event-induced graph topology rather than irrelevant frames.
- EventVAD comprises dynamic graph construction, graph attention propagation, statistical boundary detection, and event-centric MLLM scoring.These four modules establish spatiotemporal correlations, refine node features, identify event transitions, and evaluate anomalies within semantic event units.
- 3.3 Statistical Boundary Detection: Statistical boundary detection combines feature-space amplitude and directional discontinuities, smooths noise with a Savitzky-Golay filter, and thresholds signal ratios using median absolute deviation.With k=3, boundaries satisfy r_i > M, and adjacent boundary points are merged to reduce over-segmentation.
- 3.4 Event-Centric Scoring: Event semantic units address the trade-off between losing context in short fragments and weakening spatial representation in long units for visual-language models.Hierarchical prompting first generates video content descriptions and then derives anomaly scores, with two-stage reasoning supporting contextual self-correction.
4 EXPERIMENT
EventVAD is evaluated on UCF-Crime and XD-Violence using benchmark comparisons, qualitative analyses, efficiency studies, and component ablations. Results indicate that event segmentation, graph attention, optical flow, and structured reasoning support accurate detection with a smaller MLLM.
- Main Results: EventVAD surpasses training-free methods on UCF-Crime, outperforms LAVAD by nearly 4%, and also exceeds unsupervised, one-class, and some weakly supervised methods.The comparison uses a 7B Video-Llama2 foundation model.
- Main Results: EventVAD’s AP and AUC on XD-Violence are almost 5% higher than training-free SOTA LAVAD, while also surpassing one-class and unsupervised methods.The result extends the comparison beyond low-resolution or noisy scenarios to clearer, higher-resolution videos.
- Visual Analysis: Graph attention strengthens adjacent-frame connections, suppresses distant pseudo-correlations, and amplifies feature contrasts for statistical event-boundary detection.The visualization analysis links these refined temporal relationships to more accurate boundary signals.
- Visual Analysis: At score threshold 2, boundary patterns distinguish normal temporal coherence from anomalous discontinuity, while the framework remains noise-elastic yet sensitive to interframe variation.The analysis highlights frames 965–975, high-variance clips around frames 1200–1300, and stable clips around frames 1750–1900.
- Qualitative Results: Event-aware segmentation helps EventVAD localize core anomalous segments and maintain detection on long, blurry, noisy, or visually subtle videos where competing methods struggle.The reported examples include videos exceeding 2,500 and 1,400 frames, although boundary start and end points can lack precision.
- Efficiency-Accuracy Trade-off: With Video-Llama2-7B, EventVAD reaches an AUC of 82.03%, 3.69% above the 13B LAVAD baseline, indicating reduced reliance on parameter scaling.The authors attribute this efficiency-accuracy trade-off to dividing long videos into coherent segments for improved LLM comprehension.
5 CONCLUSION
EventVAD is a training-free, event-aware video anomaly detection framework that segments videos to address fine-grained transitions and diverse events. On UCF-Crime and XD-Violence, it achieves SOTA training-free performance with 7B MLLMs against larger baselines.
- 5 CONCLUSION: EventVAD achieves SOTA video anomaly detection on UCF-Crime and XD-Violence in the training-free setting with 7B MLLMs.It outperforms baseline systems using 13B or larger MLLMs.
- 5 CONCLUSION: EventVAD segments videos into events to address fine-grained visual transitions and diverse events in long videos.
- 5 CONCLUSION: Optical flow matching and dynamic graph architectures improve event boundary detection and event division for accurate within-event anomaly scoring.
- 5 CONCLUSION: Future work will explore pruning and quantization for deployment in resource-constrained scenarios.