Deep learning for surface flow velocimetry

Deep learning methods have been shown to achieve state-of-the-art velocity estimation across synthetic computer vision benchmarks and particle image datasets. Images acquired in real environments however, present additional challenges such as seeding sparsity, time-varying seeding morphology, imperfect lighting, camera stability and orientation. Therefore, we evaluate the performance of deep learning based velocity estimation methods across a range of real hydrodynamic images and compare with classical methods. We employ a hydrodynamics laboratory dataset featuring a variety of flow types and two open-source aerial river footage datasets from field campaigns. Our investigation explores three deep learning approaches which utilise different operating principles; recurrent all-pairs-field transforms (RAFT), a physics-informed approach and an unsupervised learning approach (UnLiteFlowNet-PIV). Additionally, we demonstrate the applicability of the unsupervised method for environmental flow velocimetry, where ground truth data sources are unavailable for supervised model training.