Comparative Evaluation of YOLO Models for Real-Time Detection of Multiple Construction Resources Using Single- and Multi-Source Data

The evolution of technology facilitates efficient construction-site monitoring, hence improving the assessment of project performance and resource utilization. You Only Look Once (YOLO) algorithms are employed for real-time object detection; however, their applicability for multiple construction resources is limited, their performance requires enhancement, and their tendency to overfit necessitates further investigation. This paper uses recent YOLO algorithms to develop real-time object detection models for recognizing three construction resource categories. This study evaluates newly established YOLO algorithms and conducts a cross-source generalization analysis utilizing single- and multi-source datasets to boost model adaptability across various construction environments. Three different datasets of construction resource images were collected, preprocessed, and compiled. Six YOLO models (YOLOv8–12 and YOLO26) were trained, validated, and tested for accuracy, overfitting, and generalizability. The developed models demonstrate exceptional performance in real-time detection of three construction resources, with machine detection being the most efficient. The leading models, YOLO26 and YOLOv9, have mean Average Precision (mAP) scores of 0.88 and 0.87, respectively. The preliminary findings indicate that multi-source YOLO models outperform single-source ones, exhibiting superior generalization with an approximate mAP improvement of 23%, particularly when tested on data from environments and distributions different from the training dataset. This research advances the application of innovative technologies for effective resource and construction-site management.