ABSTRACT
In the prefabrication construction site, various precast components undergo the process consisting of hoist and installation. This process contains abundant valuable information, such as components' installation time, which contribute to the project management. Therefore, aiming to extract and collect precast components temporal and locational information, this paper proposes a vision-based framework that utilizes surveillance construction site videos to detect and track the hoist and installation process of precast wall, one of the most important and widely components in prefabricated construction. This framework is mainly composed of two advanced computer vision algorithms, Mask R-CNN and DeepSORT. The Mask R-CNN realizes the object detection and instance segmentation while the DeepSORT realizes the multiple object tracking. As a result, precast walls' contours can be segmented, and walls' pixel locations as the frames progress can be obtained. The methods in framework are evaluated respectively, and the demonstration on a real project is conducted to prove the proposed framework's feasibility and efficiency. The automatic detection and tracking of precast walls realized by this paper contributes to capture and collect temporal and locational information of elements in the construction site, which lays foundation for further status judgement and analysis on elements in future work.
