Optical Flow Based Disparity Estimation Algorithm And FPGA Oriented Model Design

Optical flow estimation is one of the most important and basic researches in the field of com-puter vision.It is widely used in visual monitoring,robot navigation,image super resolution,hydrodynamics measurement,binocular disparity estimation and other tasks.Although op-tical flow algorithm is widely used,its high computational complexity and poor real-time performance greatly limit its practicability,especially in embedded applications.In the dis-parity estimation problem,the traditional disparity estimation methods based on local match-ing has a relatively low matching accuracy and relies heavily on texture features.Due to its global smoothness,the variational optical flow of global optimization can give reasonable estimation value in weak texture region.However,when the variational optimization method is directly applied to the disparity estimation task,the edge blur and details disappear due to insufficient constraints.Aiming at the problem of edge blurring and detail disappearing when variational optical flow algorithm is applied to disparity estimation,in this paper,the principle of the varia-tional optical flow algorithm is deeply studied,the causes of edge blur are analyzed,and an improved disparity estimation method based on variational optimization is realized by using the edge retention characteristics of guided image filtering.Aiming at the problems of high complexity and limited real-time performance of optical flow algorithm,this paper designs and researches the hardware implementation architecture of the mainstream optical flow estimation algorithm,and selects FPGA as the hardware platform for algorithm acceleration.The optical flow algorithm based on matching and the optical flow algorithm based on CNN are not suitable for the direct implementation by FPGA,while the optical flow algorithm based on variational optimization can be solved by numerical cal-culation method.Therefore,this paper designs the FPGA accelerated structure based on the idea of variational optimization algorithm.However,even if the variational optical flow algorithm is adopted,there are still many problems such as large computation amount and complex hardware structure of the algorithm.Aiming at these problems,this paper adopts the hardware design method based on Simulink model,and divides the variational optical flow calculation process into several stages: pretreatment,gradient calculation,linear equa-tion construction,SOR iteration solution,and designs the corresponding Simulink model for each stage respectively.Thanks to the use of code generation tools,model visualization simulation tools and model optimization tools provided by Simulink platform,the design im-plementation and verification are completed in a short period,and the algorithm is iterated to the hardware structure for many times in this paper.And the difficulties of realization,debug-ging and optimization in the hardware implementation of variational optical flow algorithm are solved.The hardware structure designed in this paper has completed the back-end design and tape verification of the corresponding chip,which will be applied to the computer vision process-ing tasks of smart phones.