An End-to-End Human Abnormal Behavior Recognition Framework for Crowds With Mentally Disordered Individuals.

Area of Science:

• Computer Science
• Artificial Intelligence
• Behavioral Science

Background:

• Abnormal and violent behaviors by individuals with mental disorders pose risks in public spaces.
• Current motion recognition technologies face challenges in accurately detecting abnormal human behavior, particularly in this population.
• Visual surveillance systems are essential for monitoring but require advanced detection capabilities.

Purpose of the Study:

• To propose an end-to-end framework for detecting abnormal behavior using Graph Convolutional Networks (GCN) and 3D Convolutional Neural Networks (3DCNN).
• To enhance the accuracy of abnormal behavior detection by modeling clip similarities and correcting noisy labels.
• To specifically improve the detection of violent behaviors in individuals with mental disorders.

Main Methods:

• A one-class classifier was trained for feature extraction and abnormality score estimation.
• Graph Convolutional Network (GCN) was employed to model video clip similarities and refine labels.
• 3D Convolutional Neural Network (3DCNN) was utilized for spatiotemporal feature extraction and behavior classification.

Main Results:

• The proposed framework demonstrated improved performance in detecting abnormal behaviors.
• GCN effectively corrected noisy labels by modeling similarities between video clips.
• On the UCF-Crime dataset, the classification accuracy reached 37.9%, surpassing existing state-of-the-art methods.

Conclusions:

• The integrated GCN and 3DCNN framework offers a promising approach for abnormal behavior detection.
• This method shows significant potential for enhancing public safety through improved surveillance.
• Further research can build upon this framework for more nuanced detection of specific abnormal behaviors.