UHPose-VAD: Unsupervised Video Anomaly Detection via Pose-Graph Learning and Normalizing Flow.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Unsupervised video anomaly detection (VAD) identifies unusual events in unlabeled videos.
• Current VAD methods often neglect fine-grained human pose dynamics, limiting localized anomaly detection.
• RGB frame-based methods struggle with lighting variations and capturing precise body structure over time.

Purpose of the Study:

• To propose UHPose-VAD, a novel unsupervised framework for human-centric video anomaly detection.
• To enhance anomaly detection by integrating human pose dynamics with normalizing flow in a graph-based probabilistic model.
• To improve robustness and interpretability in detecting localized anomalies like falls or assaults.

Main Methods:

• Extracts human pose keypoints and normalizing flow features.
• Employs a graph convolutional network with adaptive connectivity to model spatiotemporal relationships.
• Utilizes a Gaussian Mixture Model to learn the manifold of normal motion patterns in a latent space.

Main Results:

• UHPose-VAD achieves state-of-the-art performance in unsupervised VAD.
• Demonstrated high AUC scores of 86.1% on ShanghaiTech and 69.4% on UBnormal datasets.
• The framework effectively reasons about spatiotemporal joint relationships for robust anomaly detection.

Conclusions:

• UHPose-VAD offers a robust and interpretable solution for unsupervised human-centric video anomaly detection.
• The integration of pose dynamics and graph-based probabilistic modeling significantly improves anomaly detection accuracy.
• This approach addresses limitations of existing methods by focusing on precise spatiotemporal human motion patterns.