Abnormal Detection in Big Data Video with an Improved Autoencoder.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• The proliferation of video surveillance generates vast amounts of data, necessitating efficient methods for automatic anomaly detection.
• Deep autoencoders are commonly used for anomaly detection based on reconstruction errors, but can sometimes reconstruct anomalies, leading to missed detections.
• Existing autoencoder methods struggle with accurately identifying anomalies within large-scale video datasets.

Purpose of the Study:

• To develop an enhanced autoencoder model that improves the accuracy of anomaly detection in large-scale video data.
• To address the limitation of standard autoencoders in reconstructing anomalous data, which can lead to false negatives.
• To propose a novel memory-augmented autoencoder (Memory AE) for more robust video anomaly detection.

Main Methods:

• The proposed method integrates a memory module with a deep autoencoder, termed Memory AE.
• During training, the memory module learns to store representations of normal data.
• In testing, the Memory AE uses these learned memory items to reconstruct input data, amplifying reconstruction errors for anomalies.

Main Results:

• The Memory AE demonstrated improved performance in detecting anomalies compared to standard autoencoder approaches.
• Experiments conducted on the Avenue and ShanghaiTech datasets validated the effectiveness of the Memory AE method.
• The method successfully enhanced reconstruction errors for abnormal samples, leading to more reliable anomaly detection.

Conclusions:

• The memory-augmented autoencoder (Memory AE) offers a significant advancement in automatic video anomaly detection.
• By leveraging a memory module, the Memory AE effectively distinguishes normal from abnormal patterns in surveillance data.
• The proposed approach provides a more robust solution for big data anomaly detection in large-scale video surveillance systems.