Research On Binocular Vision System And FPGA Implementation

With the development of CMOS/CCD sensing technology and computer technology,machine vision has penetrated into our daily lives.In the deep learning,machine learning and artificial intelligence and other aspects have been rapidly developed,from the mobile phone beauty APP facial recognition function,Alipay's face payment,to smart driving,JD's logistics robots,etc.are using computer vision technology.Binocular Stereo Vision technology is a key subject in the field of computer vision research.It mainly uses two cameras to obtain two images from different positions based on the principle of parallax,and obtains parallax images by calculating the positional deviation of the corresponding points of the two images..Binocular vision obtains binocular parallax by simulating human eyes,and produces perception of image depth.The corresponding matching algorithm is used to determine the pixel point deviations of the two images,which can be widely used in 3D reconstruction,robot navigation,virtual reality and other fields.This dissertation compares the advantages and disadvantages of various stereo matching algorithms.Under the principle of ensuring both real-time and matching accuracy,the matching algorithm based on the improved Census transformation is selected to obtain experimental results.For the shortcomings of large amount of computation and complex calculation process in the matching algorithm,the traditional software platform can not meet its requirements.Field Programmable Gate Arrays(FPGA)are used as the hardware implementation environment platform of this thesis,Verilog HDL hardware description language to implement the functions of related modules.The entire system is divided into FPGA-based binocular image acquisition,video encoding and decoding,image data buffering,image preprocessing algorithms,and Census transformation and other related functional modules.Taking full advantage of the features of FPGA parallel processing and pipeline operations,the research of binocular vision is accomplished using a combination of hardware and software processing.In the study of image acquisition,in order to obtain more image information,this paper improves on the basis of the Demosaicing algorithm,and the processed image details are more clear.Perform preprocessing such as descaling and filtering on the initial image to prepare for subsequent matching.In the Census transformation,an improved Census transformation matching algorithm with adaptive weights is proposed for the defect of poor matching precision in discontinuous regions in stereo matching algorithm.In hardware implementation,FPGA is used to build the whole framework.This topic chooses EP4CE15F17C8 chip from Altera as processor,Quartus Prime15.1 as development tool and Modelsim 10.4b as simulation tool.The OmniVision OV5640 was used as a calibration binocular camera to capture images and two images with an output resolution of 1920*1080@60 FPS were tested.Finally,the advantages of FPGA pipeline design are fully utilized to reduce the use of logic resources and improve the matching accuracy and processing speed.