Human-independent activity recognition of construction worker

With recent advancements in sensor and data analysis technology, multiple research on worker activity recognition through wearable sensors have been conducted to solve worker safety and productivity problem at construction sites. However, most rely on pre-trained models which require re-training of each worker to take into account differences between workers. To alleviate this limitation, we propose a human-independent model that can adapt to differences in workers. Our model uses variational-denoising autoencoder with soft parameter sharing to extract common features in different construction activities, achieving 78.64% accuracy which is higher than existing benchmark models.