Deep learning-based particle tracking velocimetry (PTV) for spherical and non-spherical particles: Application to sediment transport

Particle-resolved measurements in physical experiments are crucial to the study of sediment transport. Conventional particle tracking velocimetry (PTV) relies on template matching of simple shapes via intensity thresholding and nearest-neighbor linking. However, it fails in dense particle systems where particles overlap, rotate, and vary in their apparent shapes. Here, we propose a novel computer vision-based PTV combining the YOLO algorithm for arbitrary object detection and the Kalman filter for robust data association under occlusion and noise. The framework captures particle-scale kinematics and leads to accurate measurements of the flow-depth profiles of the particle shear rate (γ ̇), rotational velocity (ω), and granular temperature (T), under various transport conditions, for idealized and natural particle shapes. Both T and ω show a clear dependence on γ ̇, consistent with predictions from granular flow theory. Moreover, comparing to spheres, irregular particles exhibit steeper T–γ ̇ and ω–γ ̇ relationships, indicating greater collisional agitation, rotation, and flow energy dissipation. These granular shape effects are more significant as the Shields number is increased, highlighting the microscopic control of particle shape on bedload dynamics. The proposed PTV technique and its potential applications are expected to provide more inslights into the fundamental mechanics of sediment transport in future research.