Robust skeleton-based AI for automatic multi-person fall detection on construction sites with occlusions

Abstract

Rapid and accurate automatic fall detection is essential for improving worker safety and reducing the severity of fall-related incidents on construction sites. To address the challenges of real-time detection in complex and obstructed construction environments, this paper develops a specialized dataset for fall scenarios and introduces a skeleton-based AI model called YOSAP-LSTM. This model integrates YOLOv8 for human detection, SORT and AlphaPose for precise tracking of human keypoints, and a 1D CNN-LSTM for classifying falls versus non-falls. This approach achieves an impressive accuracy of 98.66 % (sensitivity: 97.32 %; specificity: 99.10 %), outperforming current fall detection algorithms while maintaining high accuracy under occlusions. Deployed on an edge device (NVIDIA Jetson Xavier NX), the system runs at 6.44 fps, meeting real-time requirements for portable applications. The YOSAP-LSTM model is both robust and practical, offering significant potential for real-world use in construction by enhancing worker safety through timely fall detection in challenging environments.