Low productivity in the construction industry compared with other sectors is a longstanding concern. Traditionally, managers have relied on manual sampling of worker activities by monitoring task types and durations to identify and address productivity obstacles. However, this method is labor-intensive, error-prone, and inadequate for the dynamic demands of site management. Previous studies have highlighted the potential of utilizing wristbands equipped with inertial measurement units (IMUs) to automate activity recognition and sampling processes; however, most of these investigations have been confined to controlled laboratory settings. In this study, several commonly performed construction tasks, including those extensively researched, such as tiling, masonry, and painting, as well as less frequently studied activities like manual excavation and wall demolition, were examined under real-world construction site conditions. To enhance practical relevance for site managers, workers’ activities are categorized at two levels: first, based on their overall impact on productivity using the traditional activity sampling taxonomy, which classifies activities into direct work, indirect work, and ineffective activities; and second, by their specific activity context. A variety of hybrid algorithms that integrate convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for activity recognition were employed. Among the developed models, the hybrid convolutional neural network combined with a bidirectional gated recurrent unit (CNN-BiGRU) demonstrated the highest performance, achieving classification accuracies of 84.0% at the first level and 79.5% at the second level. These findings validate the effectiveness of using a single wristband equipped with a gyroscope and accelerometer on active construction sites. However, classification accuracy was unsatisfactory for subcategories of indirect work. This study discusses the limitations of implementing such a structure in actual construction sites, offers suggestions that provide new insights for future research, and provides the groundwork for broader practical implementation.
Monfared et al. (Thu,) studied this question.