A deep learning-enhanced PTV framework for simultaneous translational and rotational tracking of rod-like particles in non-Newtonian fluids
Open Access
- 1 October 2026
- journal article
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (19) , 195108
Abstract
This study presents a deep learning-enhanced particle tracking velocimetry framework for tracking rod-like particles in a translucent non-Newtonian fluid. The framework addresses conditions under which conventional image-processing techniques fail: particle widths approaching the resolution limit of the imaging system ( ∼ 2 pixels), spatially non-uniform illumination, and progressive inter-particle occlusion. It integrates YOLOv11 instance segmentation with a customized α – β – γ predictive filter to reconstruct the simultaneous translational and rotational trajectories of individual steel fibers in a carbopol yield-stress gel. Segmentation performance was evaluated against Hough transform and SAM baselines: YOLOv11 achieved a precision of 94.7%, a recall of 98.2% and an Intersection over Union of 84.8%, exceeding the Hough transform by more than 90 percentage points in precision. Tracking performance across fiber loadings from 25 to 800 fibers was concentration-dependent, with mean trajectory lengths decreasing from 198 to 68 frames and instantaneous detection ratios ranging from 0.64 to 0.32. Continuous long-duration trajectories are routinely recovered up to approximately 200 fibers, where 5.7%–18.8% of tracks span at least 500 of the 600 frames; above 400 fibers this falls below 2%, and the framework delivers fully resolved Eulerian velocity fields while Lagrangian analysis is restricted to selected trajectories. Regional kinematic analysis resolves statistically distinct velocity, angular velocity and orientation distributions across three geometric zones: the vertical inlet, the 90 ∘ bend and the horizontal outlet. Across the full concentration range tested, the measurements capture a systematic fiber reorientation from predominantly vertical ( 44 ∘ – 57 ∘ ) to predominantly horizontal ( 7 ∘ – 31 ∘ ) alignment. The resulting particle-scale dataset resolves the coupled translational, rotational and orientation dynamics of individual fibers in non-Newtonian suspensions. It advances the quantitative characterization of fiber kinematics in complex multiphase flows and constitutes a direct validation resource for CFD-DEM simulation frameworks of fiber transport. The dominant measurement uncertainties, centroid localization error and orientation-dependent detection bias, were identified and quantified.Keywords
Funding Information
- Fondo Nacional de Desarrollo Científico y Tecnológico (1250340)
This publication has 44 references indexed in Scilit:
- Vision-based tracking of orientation and velocity for non-spherical particlesExperiments in Fluids, 2025
- Multi-object tracking review: retrospective and emerging trendArtificial Intelligence Review, 2025
- Complete solid-body rotation rate measurements of micro-plastic curved fibers in turbulenceExperiments in Fluids, 2025
- Computer Vision – ECCV 20222022
- Particle Image Velocimetry2018
- U-Net: Convolutional Networks for Biomedical Image Segmentation2015
- The Pascal Visual Object Classes Challenge: A RetrospectiveInternational Journal of Computer Vision, 2014
- Methods for Cell and Particle Tracking2012
- Particle tracking velocimetry in three-dimensional flowsExperiments in Fluids, 1993
- Digital particle image velocimetryExperiments in Fluids, 1991