Unmanned Aerial Systems and Deep Learning for Safety and Health Activity Monitoring on Construction Sites.

Area of Science:

• Construction Safety
• Artificial Intelligence
• Computer Vision

Background:

• Construction is a hazardous industry requiring constant monitoring of worker safety and conditions.
• Integrating AI and sensor technologies can improve safety management and hazard analysis.
• Proactive safety measures are crucial for preventing accidents and protecting workers.

Purpose of the Study:

• To develop and validate a framework using unmanned aerial systems (UASs) and deep learning (DL) for construction safety analysis.
• To enhance the collection and analysis of safety activity metrics for improved construction safety performance.
• To compare the effectiveness of different DL algorithms for safety-related object detection.

Main Methods:

• A framework was developed utilizing UASs for image data collection on active construction sites.
• Deep learning (DL) algorithms, specifically Faster R-CNN and YOLOv3, were employed for object detection.
• A dataset of 7041 images was used, with a 75/25 training and testing split, focusing on safety hardhat detection.

Main Results:

• Faster R-CNN achieved a precision of 93.1% for safety hardhat detection.
• YOLOv3 achieved a precision of 89.8% for safety hardhat detection.
• The study validated the framework through a pilot case study on construction sites.

Conclusions:

• The findings demonstrate the significant impact and potential benefits of using UASs and DL in computer vision for construction safety management.
• The developed framework offers a robust strategy for analyzing safety data and protecting workers.
• Faster R-CNN proved more effective than YOLOv3 in this specific application of safety hardhat detection.