Gaussian Process Regression-Based Video Anomaly Detection and Localization With Hierarchical Feature Representation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Anomaly detection research often prioritizes local anomalies.
• Global anomalies, involving complex event interactions, are less explored but crucial for understanding events like accidents.
• Existing methods may struggle with noisy data and sparse features.

Purpose of the Study:

• To develop a hierarchical framework for simultaneous local and global anomaly detection.
• To leverage Gaussian Process Regression (GPR) for modeling spatio-temporal interactions.
• To introduce a novel method for identifying anomalies in video data.

Main Methods:

• Utilizing hierarchical feature representation and Gaussian Process Regression (GPR).
• Modeling frequent geometric relations of sparse spatio-temporal interest points (STIPs) as interaction templates.
• Developing a novel inference method to compute interaction likelihood and integrating local scores into global anomaly masks.

Main Results:

• The proposed method successfully detects both local and global anomalies.
• Gaussian Process Regression (GPR) is applied for the first time to model STIP relationships in anomaly detection.
• The framework demonstrates superior performance compared to state-of-the-art methods on four benchmark datasets.
• Achieved lower computational burden than existing approaches.

Conclusions:

• The hierarchical framework effectively detects complex, global anomalies by analyzing interactions.
• GPR provides a robust and non-parametric approach for modeling spatio-temporal relationships in anomaly detection.
• The method offers a computationally efficient and high-performing solution for video anomaly detection.