A multi-memory-augmented network with a curvy metric method for video anomaly detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection identifies events deviating from normal patterns.
• Existing methods often use Euclidean metrics, which are suboptimal for high-dimensional video data.
• High dimensionality can cause Euclidean distances to converge, hindering accurate anomaly detection.

Purpose of the Study:

• To propose a novel curvy metric to address the shortcomings of the Euclidean metric in video anomaly detection.
• To introduce a Multi-Memory-Augmented dual-flow network incorporating this new metric.
• To improve the robustness and accuracy of detecting anomalous events in videos.

Main Methods:

• Developed a new curvy metric, distinct from the Euclidean metric, for measuring differences in high-dimensional data.
• Proposed a Multi-Memory-Augmented dual-flow network that explicitly models the diversity of normal patterns.
• Integrated multiple memory units into the dual-flow network to enhance representation capacity for normal patterns.

Main Results:

• The proposed curvy metric is independent and can be integrated into existing Euclidean metric-based neural networks.
• Comparative experiments demonstrate the effectiveness of the curvy metric over the Euclidean metric.
• The Multi-Memory-Augmented dual-flow network shows robustness to normal event diversity and sensitivity to abnormal events across five datasets.

Conclusions:

• The novel curvy metric effectively overcomes Euclidean metric limitations in high-dimensional video data.
• The Multi-Memory-Augmented dual-flow network provides a robust and accurate solution for video anomaly detection.
• The proposed method demonstrates significant improvements in detecting anomalies while handling diverse normal patterns.