VIBES: Bayesian-Guided VLM for Expressway Anomaly Detection
The recently introduced framework, VIBES, leverages Vision-Language Models and Bayesian inference to identify anomalies in surveillance footage from expressways, focusing on distant targets exhibiting slight abnormal movements. To tackle issues of attention dilution and significant computational demands, the system incorporates an online Bayesian inference module that persistently analyzes vehicle paths, allowing for real-time adjustments to the probabilistic thresholds of typical driving patterns. This mechanism acts as an asynchronous trigger for accurate anomaly detection. The findings of this research are available on arXiv under ID 2604.23724.
Key facts
- VIBES is an asynchronous collaborative framework for anomaly detection.
- It uses Vision-Language Models guided by Bayesian inference.
- Focuses on far-field targets with subtle abnormal vehicle motions.
- Online Bayesian inference module evaluates vehicle trajectories.
- Dynamically updates probabilistic boundaries of normal driving behaviors.
- Serves as an asynchronous trigger for precise anomaly localization.
- Published on arXiv with ID 2604.23724.
- Addresses attention dilution and high computational costs of global frame processing.
Entities
Institutions
- arXiv