This paper presents a novel image synthesis methodology for construction site worker detection. Using a generative AI platform called Midjourney, we generate 12,000 synthetic images with 3,000 different prompts, which are then manually labeled and used as a DNN training dataset. When evaluated on a real construction image dataset, we achieve an average precision (AP) of 0.937 at an IoU threshold of 0.5 and an AP of 0.642 between 0.5 and 0.95. On the synthetic dataset, we achieve high performance with APs of 0.994 and 0.919, respectively. This demonstrates both the potential and limitations of generative AI to address the lack of DNN training data.
