An informative dual ForkNet for video anomaly detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Autoencoders are used for video anomaly detection by learning normal data representations.
• Traditional autoencoders can suffer from information loss during data reconstruction.
• Identifying abnormal events in videos requires robust and informative feature representations.

Purpose of the Study:

• To develop a novel dual ForkNet autoencoder architecture for enhanced video anomaly detection.
• To mitigate information loss in autoencoder-based anomaly detection using Informetrics Recalibration (IR).
• To improve the differentiation between normal and abnormal events through adaptive feature recalibration.

Main Methods:

• Exploration of a dual ForkNet architecture for dissociating and processing spatio-temporal representations.
• Introduction of Informetrics Recalibration (IR) to adaptively recalibrate latent features by modeling encoder-decoder similarity.
• Integration of a Secondary Encoder (SE) to refine latent feature representations.
• Utilizing ResNet blocks for a simplified and robust model architecture.

Main Results:

• The proposed dual ForkNet with IR and SE demonstrates superior performance on five public benchmarks.
• The model effectively minimizes information loss, retaining crucial semantic information for anomaly detection.
• Achieved state-of-the-art results compared to existing video anomaly detection architectures.

Conclusions:

• The novel dual ForkNet architecture with Informetrics Recalibration offers an effective solution for video anomaly detection.
• The proposed methods enhance the model's ability to differentiate between normal and abnormal events.
• The framework is robust, easy to train, and achieves state-of-the-art performance.