Region Anomaly Detection via Spatial and Semantic Attributed Graph in Human Monitoring

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Current deep learning methods for anomaly detection in human monitoring focus on single or paired regions.
• These methods struggle to capture complex interactions among multiple regions involving humans and objects.
• Existing approaches lack the ability to detect anomalies that span more than two regions.

Purpose of the Study:

• To propose a novel graph-based deep framework for detecting anomalous image regions in human monitoring.
• To overcome the limitations of existing methods in capturing multi-region interactions and complex contextual anomalies.
• To develop a model capable of understanding spatial and semantic relationships between image regions for enhanced anomaly detection.

Main Methods:

• A spatial and semantic attributed graph is proposed to model region context, incorporating spatial relations and semantic similarities.
• A Spatial and Semantic Graph Auto-Encoder (SSGAE) is developed to reconstruct this graph and detect abnormal nodes (regions).
• The graph represents regions as nodes and their contextual relationships (spatial and semantic) as edges.

Main Results:

• The SSGAE model demonstrated improved performance on three real-world datasets.
• AUC scores showed significant improvements compared to state-of-the-art baseline methods.
• Specific AUC score increases ranged from 0.79 to 0.83, 0.83 to 0.87, and 0.91 to 0.93.

Conclusions:

• The proposed SSGAE framework effectively detects anomalous image regions by leveraging spatial and semantic contextual information.
• The graph-based approach enhances the detection of complex anomalies involving multiple interacting regions.
• SSGAE offers a promising advancement in automated human monitoring and anomaly detection systems.