River Surface Velocimetry Based on Virtual River Dataset and Modulated-RAFT

Abstract

In recent years, there has been a surge of interest in image-based velocimetry methods for river surface velocity (RSV) estimation due to their efficiency and accuracy, including large-scale particle image velocimetry (LSPIV), space-time image velocimetry (STIV) and optical flow velocimetry (OFV). Among these methods, OFV-based methods have received significant attention owing to their high field resolution and low tracer requirements. Deep optical flow estimation (DOFE), which is a powerful approach for accurate and efficient estimation of optical flow, has also been employed in OFV-based methods. However, the immeasurability of optical flow often results in the usage of irrelevant datasets for training, leading to limited generalization due to domain drift. Moreover, the high similarity of river surfaces can lead to ambiguous correlation volumes extracted by DOFE models, resulting in mismatches. To address the domain shift and mismatch challenges, we proposed a method to generate optical flow datasets and these datasets are used as the training set for the DOFE model MRAFT, which incorporates correlation volume modulation. Experiment results demonstrate that our method effectively mitigates underlying domain shift and mismatch issues, enabling accurate and robust RSV estimation under velocity ranges of 0-6.0 m/s. Our work facilitates the application of DOFE on RSV estimation and provides optical flow datasets for fine-tuning to other related researches.

Keywords

• Deep learning
• optical flow
• surface flow velocity
• virtual river datasets
• correlation volume modulation