Automatic Scaffolding Productivity Analysis through Machine Learning

University: Curtin University

Industry Partner: KAEFER Integrated Services Pty Ltd

Industry Supervisor: John Folarni

IDTC Student: Wenzheng Ying

Academic Supervisors: Dr Changzhi Wu, Professor Xiangyu Wang, Australasian Joint Research Centre for Building Information Modelling (BIM), Professor Song Wang, Department of Mathematics and Statistics

Scaffold is defined as a temporary structure erected to support access or working platforms for a work crew and materials. Scaffolds are widely used in the construction, maintenance and repair of man-made structures. For example, scaffolds are erected for the maintenance process of LNG (Liquefied Natural Gas) plants during shut-down period. Scaffolding work includes erecting, altering or dismantling a temporary structure erected to support a platform, which is a labour intensive task especially under severe heat stress in outdoor environments. Current productivity analysis method of scaffolding process still remains manual watching that extra manual work is required to record the scaffolding progress. Few research has been conducted in specific action recognition and analysis of scaffolding progress. An automated method to analyse productivity and control progress of scaffolding is believed to have a promising application in construction management.

In cooperation with KAEFER Integrated Services Pty Ltd, IDTC student Wenzheng Ying proposed an interpretation method that implemented convolutional neural networks to recognise actions of scaffolding process for productivity analysis based on onsite videos. As planned, KAEFER would facilitate the onsite data collection and scaffolding practical guidance.

Neural networks have been demonstrated as a powerful class of machine learning models for image detection and recognition, which require a huge relevant database to support the training procedure of deep learning. In this study, over 100 video clips (180 frames/video on average) of scaffolding work are planned to be captured on different construction sites and also extracted from YouTube and other online video datasets to ensure sufficient data can be utilized. Furthermore, 7 consecutive frames are taken as the input of convolutional neural networks in this research.

Besides the data collection, to realize a better performance, the architecture of convolutional neural networks is deemed to be crucial in this research as it determines the balance of computational cost and recognition accuracy. Regular iterations are essential for structure adjustment and fine tuning.

Additionally, future research would be allocated on practical productivity analysis on scaffolding. Interviews with onsite managers and scaffolders would assist in discovering an appropriate method connecting typical scaffolding actions with the productivity of scaffolders. And a specific index is supposed to be developed to reflect the efficiency of scaffolders.