AI Dangerous Behavior Classification Technique Using Height and Angle Differences in Human Joint Keypoints

This paper proposes a novel approach to classify dangerous behavior using YOLOv8 pose-based keypoint extraction. Traditional CNN-based methods provide high accuracy but require significant computational resources, making real-time processing challenging. Therefore, to solve this issue, the paper proposes the method differentiates behaviors by using keypoint height differences for front view data and keypoint angles for side view data. Experimental results show that the proposed approach achieves an average accuracy of 95.7% while running 4.11 times faster (9.06ms) than CNN-based models. This improvement enables real-time detection of dangerous behavior on edge devices, making it suitable for applications in security surveillance, industrial safety monitoring, and smart city environments. The system effectively classifies punching, kicking, and falling behaviors by analyzing extracted keypoints without the need for extensive training data. The results show that our method significantly improves detection efficiency while maintaining high accuracy. This enables efficient CCTV monitoring.