INTEGRATING DOMAIN KNOWLEDGE WITH DEEP LEARNING MODEL FOR AUTOMATED WORKER ACTIVITY CLASSIFICATION IN MOBILE WORK ZONES

Accurate classification of workers' activity is critical to ensure the safety and productivity of construction projects. Previous studies in this area are mostly focused on building construction environments. Worker activity identification and classification in mobile work zone operations is more challenging, due to more dynamic operating environments (e.g., more movements, weather, and light conditions) than building construction activities. In this study, we propose a deep learning (DL) based classification model to classify workers' activities in mobile work zones. Sensor locations are optimized for various mobile work zone operations, which helps to collect the training data more effectively and save cost. Furthermore, different from existing models, we innovatively integrate transportation and construction domain knowledge to improve classification accuracy. Three mobile work zone operations (trash pickup, crack sealing, and pothole patching) are investigated in this study. Results show that although using all sensors has the highest performance, utilizing two sensors at optimized locations achieves similar accuracy. After integrating the domain knowledge, the accuracy of the DL model is improved. The DL model trained using two sensors integrated with domain knowledge outperforms the DL model trained using three sensors without integrating domain knowledge.