Automated action recognition using an accelerometer-embedded wristband-type activity tracker

Abstract

Automated worker action recognition helps to understand the state of workers’ actions, enabling effective management of work performance in terms of productivity, safety, and health issues. A wristband equipped with an accelerometer (e.g., activity tracker) allows to collect the data related to workers’ hand activities without interfering with their ongoing work. Considering that many construction activities involve unique hand movements, the use of acceleration data from a wristband has great potential for action recognition of construction activities. In this context, the authors examine the feasibility of the wrist-worn accelerometer-embedded activity tracker for automated action recognition. Specifically, masonry work was conducted to collect acceleration data in a laboratory. The classification accuracy of four classifiers—the k-nearest neighbor, multilayer perceptron, decision tree, and multiclass support vector machine—was analyzed with different window sizes to investigate classification performance. It was found that the multiclass support vector machine with a 4-s window size showed the best accuracy (88.1%) to classify four different subtasks of masonry work. The present study makes noteworthy contributions to the current body of knowledge. First, the study allows for automatic construction action recognition using a single wrist-worn sensor without interfering with workers’ ongoing work, which can be widely deployed to construction sites. The use of a single sensor also greatly reduces the burden to carry multiple sensors while also reducing computational cost and memory. Second, influences associated with the variability of movement between subject and experience group were examined; thus, a consideration of data acquisition that reflects the characteristics of workers’ actions is suggested.