Use of Computer Vision and Object Detection to Measure Construction Productivity Using Crew Balance Charts

Experts in the construction industry identify artificial intelligence (AI) technologies as a key strategy for improving productivity. In Chile, construction productivity has stagnated over the past two decades. This study explores the use of computer vision and a machine learning (ML) algorithm to measure productivity reliably, aiming to improve processes and support data-driven decision-making.
This research uses the YOLOv5 algorithm to detect workers' body postures from video and image data. Body postures are categorized as Productive or Contributory Work based on a predefined taxonomy. The algorithm was trained using 1,500 images extracted from 74 360-degree videos captured using a GoPro camera, representing over five hours of slab formwork installation. Experimental results achieved a mean average precision (mAP 0.5) exceeding 85%.
For productivity measurement, fixed-camera recordings captured images at five-second intervals. YOLOv5 detected postures for key tasks, including: installing perimeter taping (IPT), installing plumbed props (IPP), installing supporting beams (ISB), and installing formwork panels (IFP). Results were visualized through Crew Balance Charts, comparing YOLOv5-based and manually constructed analyses. IFP exhibited the best performance results and most of detected images corresponded to Productive Work.