A Unified GAN-Based Framework for Unsupervised Video Anomaly Detection Using Optical Flow and RGB Cues.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video anomaly detection is challenging in unconstrained environments due to limited labeled data and diverse scenarios.
• Existing methods struggle with the complexity and variability of real-world video data.

Purpose of the Study:

• To develop a novel unsupervised framework for video anomaly detection.
• To effectively integrate RGB appearance and optical flow motion cues.
• To improve training stability and reconstruction quality for anomaly detection models.

Main Methods:

• A unified Generative Adversarial Network (GAN)-based architecture combining dual encoders and a GRU-attention temporal bottleneck.
• A discriminator utilizing ConvLSTM layers and residual-enhanced MLPs for temporal coherence evaluation.
• Introduction of DASLoss, a composite loss function incorporating pixel, perceptual, temporal, and feature consistency terms.

Main Results:

• Achieved 80.5% Average Precision (AP) on the XD-Violence dataset, outperforming unsupervised methods like MGAFlow and Flashback.
• Obtained an AUC of 0.92 and F1-score of 0.85 on the Hockey Fight dataset for detecting short-duration violent events.
• Attained an AUC of 0.96 on the UCSD Ped2 dataset, matching state-of-the-art performance without supervision.

Conclusions:

• The proposed unsupervised framework demonstrates effectiveness and generalizability across diverse anomaly detection tasks.
• The integration of appearance and motion with a novel GAN architecture and loss function significantly improves anomaly detection performance.
• The method provides a robust solution for video anomaly detection in unconstrained environments, even with scarce labeled data.