Development of real-time hazard detection system and domain adaptation-based training framework for construction site

Abstract

The construction industry is one of the most hazardous industries with the highest number of accidents recorded globally. This thesis proposes a real-time hazard detection system for construction sites and a domain adaptation-based training framework for customizing the object detection model. The proposed system is composed of an object detection module, an object tracking module, and an image classification module. It operates in real-time using far-field surveillance videos and can perform three functions: hardhat-wearing detection, heavy-equipment operation detection, and signal-man arrangement detection simultaneously. Field application tests were performed to validate the effectiveness of the proposed system, and the results demonstrated that the system operated as intended. In addition, this thesis explores the application of unsupervised domain adaptation techniques to customize object detection models for construction site monitoring. Three domain adaptation approaches, namely self-training, reconstruction-based, and adversarial-based approaches, were employed and evaluated in this study. An unsupervised domain adaptation-based training framework enable fully-automatic customization of object detection models in construction site monitoring. Implementing this framework is expected to significantly reduce the time and labor required for manual customization, thereby enhancing the applicability of computer vision-based construction site monitoring technologies. The proposed system and training framework have great potential in construction safety management, and the results of this study provide a valuable reference for researchers and practitioners in the field.