Research On Stereo Vision Technology Under Variant Illumination

In computer vision, the topic of stereo vision has been one of the most widelystudied and fundamental problems, and continues to be one of the most activeresearch areas. Binocular stereo matching, which directly imitates the human visualsystem, is the process of taking two images and estimating a3D model of the sceneby finding matching pixels in the images and converting their2D positions into3Ddepths.Stereo vision takes the binocular cameras as sensing devices, which supportshigh-resolution, low-powered and low-payload application for mobile robotnavigation and manipulation, as well as obstacle detection and3D model building. Ithas been widely used in mobile robots such as unmanned aerial vehicles (UAVs),autonomous underwater vehicles (AUVs), unmanned ground vehicles (UGVs). Withthe expand field of scientific exploration, mobile robot works from indoorenvironments with static illumination to outdoor environment with variantillumination.In the outdoor environment, the lighting variation is proved to be a challenge forstereo matching. Traditional stereo matching techniques, which attempt to findcorrespondences by comparing the intensity difference of the pixels between the twoimages, are based on the assumption that the pixels in the image pair whichcorrespond to points in the same scene should have same intensity under uniformillumination conditions. However, considering the lighting variation in outdoorenvironment, the assumption is hard to be satisfied. The bias or gain variations of thestereoscopic cameras, as well as the viewpoint variation between them, lead todifferent intensity of the pixels in the same scene on the image pair. As the result,mismatching rate increases under lighting variation environment. Furthermore, theaccuracy of the disparity map is decreased by changes of the texture or noiseincreasing in the images caused by weak or strong illumination.In this thesis, on the basis of current research, the key problems of binocular stereo matching under lighting variation conditions are focused. A novel sub-pixelcorner detection algorithm based on CRF is proposed to improve the accuracy of thecamera calibration. The illumination-robust performance of non-parameter transformis enhanced by improving the classic sparse Census transform. Instead of using theintensity value, the adaptive support-weight based on non-parameter value isproposed to improve the matching accuracy under illumination variation conditions.At last, an integrated platform is introduced and tested. Several effective practices aredone in outdoor environment. Main innovative aspects are discussed as follows:（1） Research on accuracy sub-pixel corner detection algorithm for cameracalibrationOn the foundation of plane based camera calibration methods，the research isfocused on the accuracy of extracting the feature points of X corners. A sub-cornerdetector is developed, which is based on corner response function (CRF) as thedistribution of the weights of center of gravity (COG). The results of cameracalibration experiments show that the proposed algorithm is more accurate and robust.（2） Research on illumination-robust non-parameter transformNon-parameter transform has been proved to be more illumination-robust. Basedon the study of classic non-parameter transform methods, an improved sparse Censustransform (ISCT) is proposed. Instead of center pixel, the mean of the pixels aboveand under the center pixel is used as the reference value, which prevents errortransform from the distortion of the center pixel caused by illumination variation.After that, a stereo matching algorithm based on ISCT is designed for dense disparitymap. DSI calculation and left/right consistency are introduced to optimize therobustness. The matching quality reaches the classic algorithms on the Middleburystereo evaluation website in static illumination. Additional evaluation is done bycomparing the results with light variation, and ISCT shows better performance inrobustness.（3） Research on adaptive support-weight stereo matching algorithmArea-based local methods assume that all pixels in a support window are fromsimilar depth in a scene so that they have similar disparities. However, the supportwindows located on depth discontinuities represent pixels from different depths, and this result in the “foreground-fattening” phenomenon. Though traditional adaptivesupport-weights method would improve the accuracy on depth discontinuities, itdepends on the intensity value of the pixel, which causes the results are sensitive toillumination changes. A novel adaptive support-weight based on non-parametertransform in RGB color space is developed. The red, green and blue channels aretransformed individually through non-parameter transform, and the strength ofgrouping by proximity is calculated by Hamming distance, which reduces thedependence of the pixel intensity. The matching accuracy of the proposed method isbetter than the traditional non-parameter area-based stereo matching algorithms on theMiddlebury stereo evaluation website. In light variation environment, the algorithmshows better performance in robustness than classic adaptive support-weightalgorithm.（4） Develop an illumination-robust stereo systemOn the basis of sub-pixel corner detection algorithm for camera calibration, ISCTnon-parameter transform and adaptive support-weight stereo matching algorithm, anillumination-robust stereo system is designed for mobile robot in illuminationvariation environment. Confidence filter and memory accessing algorithm aredesigned for optimization of system efficiency and reliability. The optimizationimplementation fulfills the requirements of dense disparity map calculation andobstacle detection. Several effective tests were done in outdoor environment underlight variation, and the system shows robust performance with accurate disparity mapcalculation and obstacle detection.