Augmented dataset for multidimensional ballast segmentation and evaluation

Ballast plays an important role as a railway track supporting structure against repeated loading. Ballast degradation results in poor drainage, lateral instability, excessive settlement, service interruptions, and safety concerns. Therefore, utilizing efficient methods for evaluating ballast fouling is important to ensure safe railroad operations. Emerging computer vision approaches have been effectively incorporated into the ballast inspection processes to enhance their accuracy and robustness. Owing to the inherent data-driven nature of deep learning approaches, the efficacy of the developed model heavily depends on the quality of the training dataset, revealing an urgent need for a comprehensive and high-quality annotated ballast aggregate dataset. In this study, a multidimensional ballast aggregate dataset has been established which contains realistic data collected from field and laboratory setups, augmented by data produced by a synthetic ballast particle generator. In the dataset, the 2D data comprised ballast images with 2D annotations (masks) for localizing the particles in the images. The 3D data comprises height maps, point clouds, and 3D annotations for particle localization. The collected data included various environmental lighting levels and fouling conditions to ensure adequate coverage and diversity in the training dataset. A 2D ballast particle segmentation model previously developed by the authors was trained using an augmented dataset and demonstrated reliable and accurate results in field ballast inspections. The proposed dataset will be utilized in future studies, including 3D ballast particle segmentation and shape completion, to aid in the inspection and investigative methods for effective ballast maintenance.