Development of Automated System for Classifying Productivity Behavior of Construction Workers Using Deep Learning

Abstract

In Japan, the integration and comprehensive understanding of data related to the working environment and productivity at construction sites remain underdeveloped. This study introduces a system that utilizes the human activity recognition method, employing accelerometers combined with deep learning techniques, to capture a detailed overview of activities performed by construction site workers. We developed a new approach for transforming accelerometer data collected from devices attached to workers’ helmets into a format suitable for image-based analysis. This data was then processed using a convolutional neural network to create a deep-learning model capable of distinguishing between different types of worker movements. The model demonstrated high accuracy, with correct classification rates of 80.0% for walking and 92.1% for forward-leaning postures—activities commonly observed at construction sites. Additionally, we established an ensemble system to enhance the final classification of productive motions. This innovative system holds the promise of enabling future quantification of on-site productivity through daily indices that reflect workers’ engagement levels.