Action Recognition Using a Wristband-Type Activity Tracker: Case Study of Masonry Work

Abstract

Given that labor is one of the most important resources in a construction project, collecting field data on workers’ activities (i.e., work sampling) is critical to understanding and managing workers’ performance for a productivity analysis. Unlike manual observation used for work sampling, automated action recognition and analysis using sensors, such as motion and image sensors, enable continuous worker monitoring and corresponding task assessment. Among diverse sensors, an accelerometer has great potential for automated action recognition due to its data richness and mobility. In this paper, we propose wrist-worn accelerometer-based action recognition with selected features and classifiers and apply it to masonry work to demonstrate its feasibility. The novelty of this approach is the use of a single affordable wrist-worn sensor, which would not interfere with workers’ ongoing work. The result shows that Multilayer Perceptron classifier can achieve about 97% of accuracy in posture classification in masonry work. The proposed approach has an immense potential to be used for non-intrusive action recognition for construction workers, which can open a door for diverse productivity analyses.