Augmented Dataset for Vision-Based Analysis of Railroad Ballast via Multi-Dimensional Data Synthesis

Abstract

Ballast serves a vital structural function in supporting railroad tracks under continuous loading. The degradation of ballast can result in issues such as inadequate drainage, lateral instability, excessive settlement, and potential service disruptions, necessitating efficient evaluation methods to ensure safe and reliable railroad operations. The incorporation of computer vision techniques into ballast inspection processes has proven effective in enhancing accuracy and robustness. Given the data-driven nature of deep learning approaches, the efficacy of these models is intrinsically linked to the quality of the training datasets, thereby emphasizing the need for a comprehensive and meticulously annotated ballast aggregate dataset. This paper presents the development of a multi-dimensional ballast aggregate dataset, constructed using empirical data collected from field and laboratory environments, supplemented with synthetic data generated by a proprietary ballast particle generator. The dataset comprises both two-dimensional (2D) data, consisting of ballast images annotated with 2D masks for particle localization, and three-dimensional (3D) data, including heightmaps, point clouds, and 3D annotations for particle localization. The data collection process encompassed various environmental lighting conditions and degradation states, ensuring extensive coverage and diversity within the training dataset. A previously developed 2D ballast particle segmentation model was trained on this augmented dataset, demonstrating high accuracy in field ballast inspections. This comprehensive database will be utilized in subsequent research to advance 3D ballast particle segmentation and shape completion, thereby facilitating enhanced inspection protocols and the development of effective ballast maintenance methodologies.

Keywords

• ballast degradation
• deep learning
• computer vision
• synthetic data
• point cloud
• dataset augmentation