ActDNN: independent and sequential learning framework for accurate construction equipment monitoring

This paper presents multi-label ActDNN, a novel neural network for activity recognition on construction sites, addressing limitations in vision-based methods reliant on large structured datasets. ActDNN facilitates robust multi-label activity recognition through independent learning and sequential learning. In independent learning, the network was trained and tested on an independent set of frames, achieving an accuracy of 99.82%. In sequential learning, sequential information was utilised to predict the sequential activities of an excavator and two trucks, achieving prediction accuracies of 97.79%, 89.67%, and 86.48%, respectively. This study enhances vision-based methods for automating sequential activity and productivity analysis, offering scalable and efficient construction equipment monitoring.