Research On River Surface Velocity Estimation Based On Optical Flow Calculation

Hydrological monitoring has always been the foundation and crux of hydrological work.Through effective monitoring,acquisition,research,analysis and forecasting of hydrological data and information,it is beneficial to the protection of water resources,the development of water conservancy and flood control.The current river velocity measurement technology mainly includes three aspects: contact velocimeter,non-contact Doppler principle and non-contact image methods.This article aims at the current demand for visualized velocity measurement and quantitative description of river flow for hydrological forecasting,using optical flow as the premise,a monitoring video river surface velocity measurement method based on optical flow calculation is proposed.The main research content of this thesis includes the following work:(1).Inspired from the traditional rigid body motion tracking and detection,the traditional sparse optical flow algorithm is used to calculate the water surface flow velocity without auxiliary conditions,and a pyramid layered LK optical flow algorithm based on the scale and rotation invariant characteristics is proposed in this thesis.Increase the scale and rotation invariance characteristics of corner extraction to improve the robustness of optical flow motion estimation.When the algorithm realizes the estimation of the water surface velocity,the corresponding point reverse inspection verification is added to the pyramid multi-scale iterative method,to improve the detection accuracy of the follow-up tracking point.Then,for the extraction of river surface texture features,on the basis of the effective corner detection based on template gradient combination,feature detection of scale and rotation invariant characteristics is added to calculate high-quality features suitable for tracking,so that the features are at the sub-pixel level is more even in distribution.At the same time,this thesis is based on the large-scale feature movement distance calculation model to calibrate the filtering threshold,and eliminate the invalid value of feature point movement.(2).Visualization of two-dimensional flow field based on point spread.In order to more accurately estimate the movement direction and flow velocity of the river surface,the dense Farne Back optical flow algorithm is used to calculate the more accurate movement velocity field on the image pixels,and the numerically calculated velocity field is effectively validated through the two-dimensional velocity field visualization technology display.The flow velocity distribution of the river on the vertical profile is interfered by various factors.such as uneven bed surface,different water quality distribution,and floating objects on the water surface,so that the difference in water velocity from the edge of the river bank to the center of the river may be too large.This thesis proposes a point diffusion-based method.The idea is to divide the river vertically and uniformly along the measuring section,and calculate the weight ratio of each vertical section based on the large-scale characteristic movement distance model to realize the calculation of the river surface velocity.And this thesis has carried out the algorithm function verification and demonstration on the water flow simulation platform built in the laboratory.(3).Particle image velocity measurement technology combined with neural network model realizes the extraction of optical flow characteristics of river images.Optical flow deep neural network is currently mostly used in rigid body motion or quasi-rigid body motion in life scenes.In order to adapt to the motion estimation measurement work of river-type scenes,this thesis proposes a FlownetS optical flow convolutional neural network model based on the combination of particle images Velocity measurement technology.Using artificially generated particle image flow field data as a model training test data set for testing and evaluation of simulation experimental data and real experimental data.