Classification of construction trade and quantification of work efficiency using posture recognitions and deep neural networks

Abstract

In construction, the planners always estimate the work productivity and work efficiency of limited workers on site for delivering the projects on time. Time study is traditionally used for analysing the worker' performance to derive the productivity and efficiency. However, this manual method is too tedious and time-consuming. In Hong Kong, video cameras are always installed on construction sites to record site operations and workers' workflows. Site videos are always archived and provided solid evidence in case of construction site accidents. The videos captured tons of data which can be of better used for benchmarking worker productivity and work efficiency. Although the research endeavors advanced object recognition technology to automate the productivity analysis using modern computers, its applications of productivity analysis are highly limited. In this research, we proposed a novel approach for video-based productivity analysis by combining the use of posture recognition and deep neural networks. First, a prediction model using deep neural network and transfer learning is calibrated. Next, a knowledge base of posture information of specialty trades is developed by capturing actual postures of known specialty trades. Then, driven by the knowledge base, the workers of unknown trades are smartly classified by the computers using posture recognitions (instead of face recognition) and deep neural networks such that the work efficiency of particular trades can be determined. To demonstrate the application of the proposed method, a practical case study of a housing project in Hong Kong is used. The scope of the study is limited to one trade of steel bar bender and fixer. The computational effectiveness of using the new approach is reported. In conclusion, this new approach outperforms the traditional one in terms of result reliability and time efficiency.