Cross-Domain Video Anomaly Detection via Synthetic Data Superimposition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing approaches to video anomaly detection (VAD) often rely on same-domain training, which is limited by the difficulty of collecting frame-level annotations for anomaly videos across different domains, leading to inefficiencies in cross-domain VAD.
Innovation Solution
The method involves generating synthetic video clips by extracting depictions of persons and their movements from annotated source-domain videos and superimposing them onto unannotated target-domain videos, creating labeled training data for machine learning models to perform cross-domain VAD.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If same-domain training is used for VAD, then measurement precision of anomaly detection is improved, but adaptability to different domains deteriorates
Solution Approach 1:
The patent creates synthetic cross-domain training data by copying and superimposing anomaly behaviors from source domain videos onto target domain videos. This allows the model to learn domain-invariant anomaly patterns while maintaining adaptability to different domains, resolving the contradiction between precision and adaptability.
Solution Approach 2:
The patent transforms the training data parameters by generating synthetic videos with modified domain characteristics while preserving anomaly behavior patterns. This enables the model to generalize across domains while maintaining detection precision through consistent anomaly pattern recognition.
2Adaptability or versatility
If frame-level annotations are collected across multiple domains, then adaptability to different domains is improved, but productivity of data collection deteriorates
Solution Approach 1:
Instead of collecting and annotating data across multiple domains, the patent copies annotated anomaly behaviors from a single source domain and superimposes them onto multiple target domains. This synthetic data generation approach achieves cross-domain adaptability without the productivity loss of manual multi-domain annotation.
Solution Approach 2:
The patent performs preliminary annotation work on source domain data only, then uses this pre-annotated data to generate synthetic training examples for multiple target domains. This preliminary action eliminates the need for repeated annotation efforts across domains, significantly improving data collection productivity.
3Productivity
If synthetic video generation is used, then productivity of training data generation is improved, but manufacturing precision of realistic video quality deteriorates
Solution Approach 1:
The patent copies real video frames and superimposes real annotated anomaly behaviors onto them, preserving the visual quality and realism of the original footage. This approach maintains manufacturing precision while achieving high productivity through automated synthetic generation.
Solution Approach 2:
The patent merges real video content from target domains with annotated anomaly behaviors from source domains through superimposition. This combination preserves the visual fidelity of real videos while incorporating labeled anomaly data, achieving both high quality and productivity.
Data Source
AI summary
Operations include extracting a depiction of a person and associated movement of the person from a first video clip of a first training video included in a first domain dataset. The operations further include superimposing the depiction of the person and corresponding movement into a second video clip of a second training video included in a second domain dataset to generate a third video clip. The operations also include annotating the third video clip to indicate that the movement of the person corresponds to a particular type of behavior, the annotating being based on the first video clip also being annotated to indicate that the movement of the person corresponds to the particular type of behavior. Moreover, the operations include training a machine learning model to identify the particular type of behavior using the second training video having the annotated third video clip included therewith.