Abstract
The adoption of automation technologies across various industries has significantly increased in recent years. Despite the widespread integration of robotics in many sectors, the construction industry remains predominantly reliant on manual labour. This study is motivated by the need to accurately recognise construction worker activities in labour-intensive environments, leveraging deep learning (DL) techniques to enhance operational efficiency. The primary objective is to provide a decision-support framework that mitigates productivity losses and improves time and cost efficiency through the automated detection of human activities. To this end, sensor data were collected from eleven different body locations across five construction workers, encompassing six distinct construction-related activities. Three separate recognition experiments were conducted using (i) acceleration sensor data, (ii) position sensor data, and (iii) a combined dataset comprising both acceleration and position data. Comparative analyses of the recognition performances across these modalities were undertaken. The proposed DL architecture achieved high classification accuracy by incorporating long short-term memory (LSTM) and bidirectional long-term memory (BiLSTM) layers. Notably, the model yielded accuracy rates of 98.1% and 99.6% for the acceleration-only and combined datasets, respectively. These findings underscore the efficacy of DL approaches for real-time human activity recognition in construction settings and demonstrate the potential for improving workforce management and site productivity.