An Angle-Based Dangerous Behavior Classification Edge System Implementation Using Deep Learning Pose Estimation Model

This paper proposes a system to address safety issues in public spaces by using a deep learning model to classify dangerous behaviors based on keypoint angles. The YOLOv8-Pose model is employed to extract keypoints in real-time, detecting dangerous actions such as punching, kicking, and fall down. Experimental results show that the proposed method achieves an accuracy of 95.6% with a processing speed of 8.1ms, making it approximately 4.65 times faster than traditional CNN-based methods. This efficiency allows for real-time monitoring in various settings, including parks, factories, and schools, without the need for high-performance devices or extensive datasets. By analyzing the angles between keypoints, the system significantly reduces computational load while maintaining high accuracy. The practical implementation of this system can enhance safety monitoring by providing an effective and efficient solution for detecting potential risks in large public areas. This system holds promise for improving safety management in public spaces through a cost-effective and computationally efficient approach.