Weapon operating pose detection and suspicious human activity classification using skeleton graphs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Security Studies

Background:

• Rising global concern over violent protests and armed conflict in civilian areas.
• Current surveillance methods are labor-intensive and inefficient for monitoring large crowds.
• Existing pose estimation techniques struggle with detecting weapon-related activities.

Purpose of the Study:

• To develop a comprehensive human activity recognition approach for detecting suspicious behaviors in crowds.
• To enhance crowd management strategies through automated surveillance analysis.
• To improve the accuracy of identifying specific violent actions during public disturbances.

Main Methods:

• Utilized a customized dataset and VGG-19 backbone to extract body coordinates.
• Developed a human activity recognition model based on human body skeleton graphs.
• Employed a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) with a Kalman filter for pose identification.

Main Results:

• Categorized human activities into eight classes relevant to violent clashes.
• Achieved 89.09% accuracy in real-time pose identification.
• Successfully differentiated suspicious activities (e.g., weapon handling) from regular actions.

Conclusions:

• The proposed end-to-end pipeline offers a robust solution for multiple human tracking and suspicious activity detection.
• The model effectively categorizes human activities, enabling targeted alarm triggers.
• This approach significantly improves crowd management by providing real-time insights into potentially dangerous behaviors.