Harnessing Synthetic Image Datasets for Enhanced Scene Understanding in Construction Work Zones

This study pioneers the use of synthetic image datasets, generated via the Unity game engine, to train deep learning models for construction work zone scene understanding. This innovative approach simplifies data acquisition and ensures a rich, diverse dataset that includes various construction scenarios, both hazardous and typical. We created 12,360 images with accurate bounding box annotations, ensuring high-quality, consistent data crucial for model training and validation, and effectively addressing the ambiguities often found in real-world datasets. An object detection benchmark was established using this dataset alongside eight state-of-the-art object detectors. This benchmark thoroughly evaluates the performance of these detectors on a wide range of construction site images, enabling comparisons and analyses of different models. It highlights their respective strengths and weaknesses in construction site applications. Notably, YOLOv8-L demonstrated exceptional performance, achieving a mean average precision (mAP) of 70.7% on the validation set and 69.8% on the testing set. These results underscore the efficacy of synthetic datasets in training models for complex scene understanding. This integration of synthetic and real-world imagery has the potential to revolutionize scene comprehension in construction zones, significantly enhancing safety and efficiency in the construction industry.