G29: Active Materials

Tue. March 5, 12:42 p.m. – 12:54 p.m. CST

101J

Obtaining accurate flow velocities from experimental video data plays an important role in developing and testing models of physical systems. While Particle Image Velocimetry (PIV) is a well-known tool for estimating velocities, it has significant limitations and may produce highly inaccurate results for many active matter systems. In this work, we develop an optical flow algorithm for active materials, which produces accurate velocity fields from experimental videos that are well beyond the limitations of PIV algorithms. Deep learning-based optical flow is a machine learning technique that uses deep convolutional neural networks to extract features in a pair of adjacent video frames and uses those features to estimate the inter-frame motions of individual pixels. We test the performance of the optical flow algorithm on microtubule-based active nematics. In our experiments, the microtubules were sparsely and densely labelled using two different fluorescent dyes. The sparse labels on the microtubules serve as “seeding particles” that can be used by PIV to accurately compute the flow fields, while the dense labels, which visualize 100% of the microtubule bundles, pose a great challenge to PIV. Velocity ground truths were obtained by performing semi-automated particle tracking on a subset of microtubule segments that are illuminated by the sparse fluorescent labels. The evaluation shows that PIV and optical flow performed equally well using the sparse labels. When using the dense labels, optical flow produces accurate velocity fields, but PIV produces highly inaccurate, unphysical results. Since sparse labeling of seeding particles is infeasible for many systems, our results imply that the optical flow technique is much more applicable. Moreover, the optical flow implementation is simpler than PIV. Thus, we expect that optical flow can become a widely used tool for velocity estimation in a broad class of active and other soft matter systems.