3D-Convolutional Neural Network with Generative Adversarial Network and Autoencoder for Robust Anomaly Detection in...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of surveillance devices necessitates advanced methods for video anomaly detection.
• Existing machine learning approaches for video anomaly detection face challenges, particularly with limited anomaly data.

Purpose of the Study:

• To propose a novel hybrid deep learning model for effective video anomaly detection.
• To address the challenge of deficient anomaly data during model training.

Main Methods:

• A hybrid deep learning model combining a generative adversarial network (GAN) for feature extraction and a transfer learning-based anomaly detector.
• Training the video feature extractor using a GAN with limited anomaly data.
• Boosting the anomaly detector by transferring the pre-trained feature extractor.

Main Results:

• The proposed model achieved a recall of 94.4% and a precision of 86.4% on the UCSD pedestrian dataset.
• Demonstrated competitive performance compared to existing state-of-the-art methods in video anomaly detection.
• The transfer learning approach effectively leveraged the trained feature extractor for improved anomaly detection.

Conclusions:

• The hybrid deep learning model offers a robust solution for video anomaly detection, even with scarce anomaly data.
• The integration of GAN-based feature extraction and transfer learning significantly enhances detection performance.
• This approach represents a significant advancement in the field of intelligent video surveillance.