Advancing Video Anomaly Detection: A Bi-Directional Hybrid Framework for Enhanced Single- and Multi-Task Approaches

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current video anomaly detection often uses multi-task learning but with suboptimal single-task frameworks.
• Optimizing individual proxy tasks can enhance both single- and multi-task anomaly detection performance.

Purpose of the Study:

• To develop and validate an optimized single-task framework for video anomaly detection.
• To demonstrate the benefits of this framework in both standalone and multi-task learning scenarios.

Main Methods:

• A hybrid framework combining vision transformers and ConvLSTMs for middle-frame prediction is proposed.
• A bi-directional structure analyzes temporal dynamics by predicting frames forwards and backwards.
• A convolutional temporal transformer and a layer-interactive ConvLSTM bridge are utilized for enhanced prediction accuracy.

Main Results:

• The hybrid framework achieves high efficacy as a single-task approach for video anomaly detection.
• Integration of the framework into multi-task approaches also yields significant improvements.
• Experiments confirm the advantages of merging vision transformers and ConvLSTMs for robust anomaly detection.

Conclusions:

• Optimized single-task frameworks are vital for advancing video anomaly detection.
• The proposed hybrid framework effectively detects anomalies by analyzing frame prediction discrepancies.
• Combining vision transformers and ConvLSTMs offers a powerful solution for video anomaly detection.