Utilizing 360-Degree Images for Synthetic Data Generation in Construction Scenarios

Abstract

Computer vision-based safety monitoring requires machine learning models trained on generalized datasets covering various viewpoints, surface properties, and lighting conditions. However, capturing high-quality and extensive datasets for some construction scenarios is challenging at real job sites due to the risky nature of construction scenarios. Previous methods have proposed synthetic data generation techniques involving 2D background randomization with virtual objects in game-based engines. While there has been extensive work on utilizing 360-degree images for various purposes, no study has yet employed 360-degree images for generating synthetic data specifically tailored for construction sites. To improve the synthetic data generation process, this study proposes a 360-degree images-based synthetic data generation approach using Unity 3D game engine. The approach efficiently generates a sizable dataset with better dimensions and scaling, encompassing a range of camera positions with randomized lighting intensities. To check the effectiveness of our proposed method, we conducted a subjective evaluation, considering three key factors: object positioning, scaling in terms of object respective size, and the overall size of the generated dataset. The synthesized images illustrate the visual improvement in all three factors. By offering an improved data generation method for training safety-focused computer vision models, this research has the potential to significantly enhance the automation of the construction safety monitoring process, and hence, this method can bring substantial benefits to the construction industry by improving operational efficiency and reinforcing safety measures for workers