Zero-Shot Cross-Domain Video Anomaly Detection

The patent uses a generator network to create synthetic pseudo-normal frames that copy the statistical and structural properties of target domain normal data without requiring actual access to target domain videos. These synthesized frames serve as proxies for real normal data, enabling model training in a zero-shot cross-domain setting while respecting security and privacy constraints.

The patent introduces an intermediary synthesis module that generates pseudo-normal frames acting as a mediator between the source domain training data and the target domain anomaly detection task. This intermediary component bridges the domain gap without requiring direct access to target domain normal data, resolving the contradiction between detection reliability and data access feasibility.

The patent segments the anomaly detection task into distinct functional modules: a generator network for synthesizing pseudo-normal frames, a discriminator network for distinguishing real from synthetic frames, and an anomaly detection model for identifying abnormal events. This segmentation allows each component to be trained independently with specific objectives, managing overall system complexity while achieving zero-shot cross-domain detection.

The patent performs preliminary action by pre-training the generator and discriminator networks on source domain data before deploying the anomaly detection model to the target domain. This preliminary training phase creates a robust feature extraction capability and synthetic data generation ability that simplifies the subsequent zero-shot detection task, reducing the operational complexity despite the sophisticated architecture.

The patent applies local quality by extracting foreground objects from source domain videos and using them to corrupt specific local regions of pseudo-normal frames. This localized corruption approach ensures that anomalies appear in semantically meaningful locations with realistic appearances, improving synthesis quality while maintaining the ability to train without target domain abnormal data.

The patent creates composite training samples by combining pseudo-normal frames (synthesized from source domain statistics) with foreground objects (extracted from source domain videos) to generate pseudo-abnormal frames. This composite approach leverages the strengths of both synthetic and real data components, achieving adequate anomaly synthesis quality for effective training in zero-shot cross-domain settings.