CRCL: Causal Representation Consistency Learning for Anomaly Detection in Surveillance Videos

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Video Anomaly Detection (VAD) is crucial for public safety and information forensics.
• Existing unsupervised VAD methods struggle with scene changes and subtle anomalies due to overgeneralization.
• Current models crudely describe normality, lacking exploration of underlying causal relationships.

Purpose of the Study:

• To propose a novel unsupervised VAD method inspired by causality learning.
• To address limitations of existing methods in handling scene-independent biases and subtle anomalies.
• To develop a method that learns robust, causal representations of normal video patterns.

Main Methods:

• Introduced Causal Representation Consistency Learning (CRCL).
• Employed structural causal models to mine scene-robust causal variables.
• Implemented scene-debiasing learning to remove scene bias from deep representations.
• Developed causality-inspired normality learning to capture causal video normality.

Main Results:

• CRCL demonstrates superiority over conventional deep representation learning methods on benchmarks.
• The method effectively handles label-independent biases in multi-scene settings.
• CRCL maintains stable performance even with limited training data.

Conclusions:

• CRCL offers a more systematic and robust approach to unsupervised Video Anomaly Detection.
• The causality-inspired framework enhances generalization and resilience to real-world complexities.
• This method advances the field of video understanding for critical applications.