Research On The Stereo Matching Based On Small Baseline Images

Stereo matching technology is an important research direction in computer vision that is widely used in robot autonomous navigation,object recognition and tracking,aerospace photogrammetry and industrial control and detection and so on.The small baseline images are obtained under a certain base to height ratio constraint condition observation model.The small baseline image stereo matching method computes the sub-pixel level disparity of a pixel,implementing a high precision stereo matching.In the space photogrammetry,stereo matching methods can be used to obtain the disparity of a pixel in the image.The terrain elevation information of the pixel can be computed by the disparity principle.As the small baseline image has the characteristics of similar simultaneity and small visual angle for terrain objects,the small baseline stereo matching methods can effectively solve the no matching owing to the city ground object sheltering,the decline of matching accuracy caused by geometric distortion and other problems.But in the process of stereo matching for small baseline images,the decrease of baseline will generate the loss of depth accuracy.Therefore,sub-pixel level stereo matching methods suitable for small baseline images are needed for the implementation of high precision stereo matching.This paper has researched on stereo matching methods for small baseline images,respectively in matching cost computation based on non-parameter transformation,matching cost aggregation and original disparity computation based on pixel expansion,disparity optimization based on image segmentation and sub-pixel level disparity computation based on Lagrange interpolation.The paper proposes a stereo matching method for small baseline images that can implement the sub-pixel level stereo matching with high precision.Firstly,aiming at matching cost computation methods based on non-parameter transformation,the paper has researched the gray-level and gradient transformation methods for solving the problems of the decline of the accuracy rate for the distortion of middle pixel in the transform window,the distortion of pixels in continuous regions,the same transform results for different pixels.A gray-level four-moded Census transform method is proposed which can better recognize the more similar transform windows in the searching scope,andsolve the problem that it can not accurately obtain the dense disparity results for the gray-level distortion of middle pixels in the transform window.A gradient four-moded Census transform method is proposed which can solve the problem of accuracy decline for illumination distortion of pixels in continuous regions and enhance the matching precision for continuous illumination change stereo images.Combining the two kinds of transforms,the paper proposed four-moded Census transform matching cost computation method based on non-parameter transformations,which can provide the more accurate matching cost in the later cost aggregation and original disparity computation.Secondly,as for the foreground expansion in the original disparity map,a matching cost aggregation and original disparity computation method is proposed based on pixel expansion.After establishing the adaptive aggregation windows,the model of windows are established by using the sizes of aggragation windows and the relative position of datum point in the windows,implementing a pre-matching of the model of windows.In the original search range,the pre-matching can reject the model of windows which have large differences from the model of window of the pixel for matching.The mismatching rate is decreased by re-establishing the search range.A matching cost aggregation method of window regularization is proposed for obtaining the original disparity map and effectively solving the foreground expansion in the original disparity map.Thirdly,owing to quite large errors in the original disparity map comparing with the ground truth,a disparity refinement method based on Mean-shift segmentation is proposed.This method firstly detects the disparity consistency of pixels in the reference image and target image to obtain the position information of error pixels in disparity map.Then the original image is segmented using Mean-shift algorithm.By merging the neighbor segments by color similarity of model pixels and edge pixels,the disparity support domain is obtained.At last,using the disparity values of non-error pixels in disparity support domain,the disparity values of error pixels in disparity support domain are updated.The accurate disparity results of the error pixels and final dense disparity map are obtained.Comparing to the original disparity map,the final disparity map using the disparity refinement method by Mean-shift segmentation is closer to the ground truth with higher matching accurate.Finally,aiming at the loss of depth accuracy caused by the small base to height ratio in small baseline images,a sub-pixel level disparity computation method is proposed based onLagrange interpolation after obtaining the precise integer level disparity results.The method determines the interpolation samples by pixel cutting resampling and obtains the fitting curves of the corresponding pixels by Lagrange interpolation theorem.It determines the original sub-pixel level disparity of pixel for matching.Using the monotony correlation between the fitting curves of corresponding pixels and the fitting curves of pixel for matching,the pixels for matching in sub-pixel level can be positioned precisely.The method can achieve a more accurate sub-pixel level disparity result and implement a higher precise sub-pixel level stereo matching.The proposed stereo matching method for small baseline images makes progress in overcoming the mismatching of illumination distorted pixels,foreground expansion of disparity map and loss of depth accuracy.The proposed method can reach the sub-pixel level matching precision and it can be applied to urban space photogrammetry.The elevation of terrain objects can be calculated by using the results of the sub-pixel level stereo matching.