Single timestep 3D particle reconstruction at high seeding densities using neural networks

Abstract

Single timestep particle reconstruction via Triangulation or IPR [1,2] is a key component of Lagrangian particle tracking methods. Entry point for such reconstructions is the detection of 2d particle peak location within the image. As particle seeding density increases the quality of peak detection as well as triangulation decreases, giving an increased number of ghosts and less accurate positions. We present an approach using a neural network based peak detector together with an IPR parameter optimization scheme using evolutionary algorithms that provide improved performance even at high seeding densities. The peak detector itself can also be applied to simpler approaches such as direct triangulation for conventional particle tracking. We validate the approach using a synthetic test case and are able to fully solve an IPR problem for seeding densities up to 0.2 ppp for clean and 0.16 ppp for noisy imaging conditions, outperforming the current state of the art.