Dense particle tracking using a learned predictive model

Abstract

The velocity resolution for particle tracking velocimetry is limited by the ability to link multiple instances of the same particle captured over time to form trajectories. This becomes increasingly difficult as the particle speed and concentration increase. To address these concerns, we propose a data-driven approach to generate a learned predictive model capable of accurately estimating future particle behavior. The model uses the long short-term memory (LSTM) recurrent neural network architecture to predict a particle’s velocity from its past positions. The model achieves increased linking performance with a negligible increase in the computational cost. Historical trajectories demonstrating the range of expected behaviors for a given application are used to train the model and can be collected using either of two methods. Manual filtering and selection can be used to select exemplary trajectories produced by an incomplete or inadequate method. Supplemental experiments with reduced tracer concentration can also produce training data. Both methods are demonstrated through experimental validation. The ability of the learned predictor to accurately link particles at tracer concentrations and flow speeds where traditional methods fail is demonstrated using two simulated flow cases. Three experimental cases are presented to demonstrate the performance of the proposed method under challenging conditions. Each case tracks the 3D positions of particles captured using microscopic digital inline holography although the approach is generalizable and can be applied to data obtained with other 2D and 3D PTV techniques. The selected cases—turbulent channel flow, flow through a T-junction, and swimming microorganisms—demonstrate the broad applicability of the proposed method to different fields with unique challenges.

Graphic abstract