An efficient non-spherical particle tracking strategy based on deep-learning and simulation-experiment integration

Abstract

Image-based particle recognition and tracking are essential techniques for experimental investigations of granular systems. Conventional methods relying on geometric features are predominantly applied to spherical or near-spherical particle systems. The rapid development of machine-learning-based image segmentation methods has facilitated the identification of non-spherical particles. However, the laborious manual annotation required for model training limits their scalability or practical applicability. Moreover, these methods typically require substantial time and effort to rebuild algorithms or datasets when applied to particles of other shapes. To address these challenges, this study presents a novel non-spherical particle tracking strategy based on deep-learning and simulation-experiment integration. The Mask Region-based Convolutional Neural Network (Mask R-CNN) is pre-trained on the synthetic dataset generated using superquadric Discrete Element Method (DEM) simulation to establish initial detection capability. The model is subsequently fine-tuned using a small number of manually corrected experimental images to enhance the robustness against real-world noise such as inter-particle overlap and occlusion, insufficient resolution and illumination, and reflection artifacts. Finally, a Particle Tracking Velocimetry (PTV) method for non-spherical particles is developed based on masks predicted by the fine-tuned model. The AI-based PTV method is evaluated and validated in rotating drum experiments, demonstrating its adaptability and reliability across different particle shapes. The proposed strategy enables the rapid development of PTV methods for new non-spherical particles, providing a practical and generalizable solution for studying the flow behavior of non-spherical particles.