Answers are grounded in this paper and cite their sources.
Effective monitoring of construction equipment activities is essential for improving operational efficiency on construction sites. Current vision-based methods are highly dependent on large and structured datasets that are impractical to collect from dynamic construction environments. This paper introduces activity recognistion deep learning neural network (ActDNN), a scalable multilabel machine learning framework designed for accurate activity recognition and productivity monitoring in near real time. It effectively handles unstructured image data, even in challenging conditions such as occlusion, varying camera angles, and when equipment is obscured by dust and dirt. The key contribution of ActDNN lies in its integration of two learning paradigms: (1) frame-based classification through independent learning to detect multiple concurrent equipment activities; and (2) temporal activity prediction using sequential learning to model activity transitions over time. ActDNN was evaluated using 8,926 frames and a whole video in these two experimental setups. The frame-based model achieved 99.82% accuracy in multilabel activity classification, while the sequential model achieved 97.79%, 89.67%, and 86.48% accuracy for the excavator and truck activity sequences, respectively. It also achieved 78.96%, 84.58%, and 54.84% accuracy in productivity prediction and 89.60%, 91.24%, and 99.10% accuracy in utilization ratio prediction for the excavator, Truck1, and Truck2, respectively. ActDNN significantly reduces the time and effort required for activity tracking, offering a practical near real-time solution for monitoring equipment.
No takes yet. Share an insight, caveat, or question.
Verma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: