NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·Jun 9·5 min read
MemoVAD: Resource-Efficient Video Anomaly Detection via Dynamic Semantic Memory in Edge Computing Scenarios
TL;DR
Video Anomaly Detection (VAD) in surveillance struggles with high computational demands on edge devices. MemoVAD is introduced as an edge-cloud collaborative framework that integrates Vision-Language Model (VLM) semantics selectively.
✦ Why It Matters
Engineers can leverage MemoVAD to implement efficient video anomaly detection in edge computing environments.
Key Takeaways
- Utilize MemoVAD for efficient video anomaly detection in edge computing.
- Implement Uncertainty-Aware Gating to optimize cloud queries.
- Leverage Dynamic Semantic Memory for improved semantic integration.
- Test MemoVAD on UCF-Crime and XD-Violence datasets for validation.
How It Works
MemoVAD operates by processing most video data on edge devices using a lightweight detector and a Temporal Context Encoder (TCE) to capture temporal relationships. The Uncertainty-Aware Gating (UAG) policy, based on Subjective Logic, assesses uncertainty in detected anomalies and decides when to query the cloud for additional semantic insights.
The Dynamic Semantic Memory (DSM) stores previously verified semantic prototypes, allowing the edge model to adaptively enhance its detection capabilities without constant cloud reliance.