Activity classification using accelerometers and machine learning for complex construction worker activities. (March 2021)

• Record Type: Journal Article
• Authors: Sanhudo, Luís; Calvetti, Diego; Martins, João Poças; Ramos, Nuno M.M.; Mêda, Pedro; Gonçalves, Miguel Chichorro; Sousa, Hipólito
• Is Part Of: Journal of building engineering. Volume 35(2021)
• Publication Date: 2021-03
• Subjects: Construction workers -- Activity classification -- Productivity analysis -- Wearable accelerometers -- Supervised machine learning; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2020.102001
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: Automated Construction worker activity classification has the potential to not only benefit the worker performance in terms of productivity and safety, but also the overall project management and control. The activity-level knowledge and indicators that can be extracted from this process may support project decision making, aiding in project schedule adjustment, resource management, construction site control, among others. Previous works on this topic focused on the collection and classification of worker acceleration data using wearable accelerometers and supervised machine learning algorithms, respectively. However, most of these studies tend to consider small sets of activities performed in an instructed manner, which can lead to higher accuracy results than those expected in a real construction scenario. To this end, this paper builds on the results of these past studies, committing to expand this discussion by covering a larger set of complex Construction activities than the current state-of-the-art, while avoiding the need to instruct test subjects on how and when to perform each activity. As such, a Machine Learning methodology was developed to train and evaluate 13 classifiers using artificial features extracted from raw accelerometer data segments. An experimental study was carried out under the form of a realistic activity-circuit to recognise ten different activities: gearing up; hammering; masonry; painting; roughcasting; sawing; screwing; sitting; standing still; and walking; with most activities being a cluster of simpler tasks (i.e. masonry includes fetching, transporting, and laying bricks). Activities were initially separated and tested in three different activity groups, before assessing all activities together. It was found that a segment length of 6 s, with a 75% overlap, enhanced the classifier performance. Feature selection was carried out to speed the algorithm running time. A nested cross-validation approach was performed for hyperparameter tuning and classifier training and testing. User-dependent and -independent approaches (differing in whether the system must undergo an additional training phase for each new user) were evaluated. Results indicate that accelerometers can be used to create a robust system to recognise large sets of Construction worker activities automatically. The K-Nearest Neighbours and Gradient Boosting algorithms were selected according to their performances, respectively, for the user-dependent and -independent scenarios. In both cases, the classifiers showed balanced accuracies above 84% for their respective approaches and test groups. Results also indicate that a user-dependent approach using task groups provides the highest accuracy. Highlights: Acceleration-based classification of ten complex construction worker's activities. Uninstructed data acquisition with six workers and three wearable accelerometers. Evaluation of thirteen classifiers for subject-dependent and independent approaches. Results prove the potential of this method in classifying large sets of activities. Best algorithm shows an average accuracy of 93.69% in a subject-dependent approach.