Using a Trained Machine Learning Algorithm for Microparticle Tracking on an Inhomogeneous Disk Array Background

When analyzing motion of micro-scale objects on inhomogeneous backgrounds, traditional tracking methods may yield limitations in location determination and trajectory linking, relying on gap closing and other estimations to determine objects' trajectories. This research aims to improve upon particle tracking using machine learning techniques to address these challenges. In particular, these algorithms train on images of microspheres whose appearance changes with location due to the inconsistency of the image background. Using superparamagnetic microparticles transported on a grid of permalloy disks by weak, time-varying magnetic fields as an example, we compare tracking using TrackMate in ImageJ and the LodeSTAR method for machine learning detection, highlighting the strengths and limitations of each approach. By implementing the Crocker-Grier algorithm for particle linking, we observe improvements in particle detection and tracking reliability, and we discuss circumstances where these improvements would be especially advantageous.