Author(s): Saber Ansari; Colin Rennie; Elizabeth Jamieson; Ousmane Seidou; Shawn Clark

Linked Author(s): Shawn Clark, Colin Rennie

Keywords: Surface velocimetry; Optical flow; Deep learning; Hydrometric measurements

Abstract: Streamflow observation and measurement are important for various water resource engineering applications, such as hydrological and hydraulic studies. Conventional methods for measuring flow velocity, such as impeller flowmeters and acoustic technologies, are intrusive and are typically used for periodic and noncontinuous measurements and are not practical during extended extreme floods and presence of ice. Nonintrusive methods, including surface velocity radar, and image-based surface velocimetry, have gained popularity in recent years due to their non-intrusive nature, practicality, and ability to provide continuous and high-resolution temporal data. For image-based surface velocity estimation, cross-correlation and gradient based techniques are the most established and accepted processing methods. This work compares two wellknown conventional surface velocimetry techniques of Large-Scale Particle Image Velocimetry (LSPIV) and Space Time Image Velocimetry (STIV) with a recently introduced deep learning based optical flow technique called RiVQNet. RivQNet is an efficient and accurate application of Convolutional Neural Network (CNN) - based optical flow approach for surface velocimetry that can be used for instantaneous and continuous data collection using aerial and oblique imagery. The comparison has been done by processing videos captured for different flow conditions and camera angles. These results indicate that RivQNet may be more reliable in variable illumination and brightness conditions, inferior quality imagery. Results indicated that under these circumstances, RivQNet has significantly higher estimation accuracy than both LSPIV and STIV, with approximately 25% and 15% higher accuracy, respectively.

DOI: https://doi.org/10.3850/978-90-833476-1-5_iahr40wc-p1697-cd

Year: 2023