An explainable and efficient deep learning framework for video anomaly detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods excel at video anomaly detection but require extensive training data and time.
• Current models lack interpretability, hindering real-world deployment where understanding anomaly severity is crucial.

Purpose of the Study:

• To develop an efficient deep learning framework for video anomaly detection with fast deployment capabilities.
• To introduce interpretability into video anomaly detection models, explaining the reasoning behind detected anomalies.

Main Methods:

• Utilized pre-trained deep models to extract features for training a denoising autoencoder (DAE).
• Integrated SHapley Additive exPlanations (SHAP) with the autoencoder framework for model interpretability.
• Evaluated the framework on the UCSD Pedestrian datasets (Ped1 and Ped2).

Main Results:

• Achieved comparable detection performance to leading methods with significantly reduced training times (under 10 seconds).
• Demonstrated high Area Under the Curve (AUC) scores: 85.9% on UCSD Ped1 and 92.4% on UCSD Ped2.
• Successfully provided explanations for anomaly detection results in surveillance videos.

Conclusions:

• The proposed framework offers an efficient and interpretable solution for video anomaly detection.
• This approach enables faster deployment of video analysis systems in real-world scenarios.
• Combining autoencoders with SHAP provides valuable insights into anomaly detection decisions.