一种新的立体视差估计算法

为了减少误匹配率，提高PSNR，减少计算复杂度，提出了一种新的固定块立体视差估计算法．根据连续性和惟一性约束，建立了高斯模型计算先验概率，用最大后验概率准则判断最佳匹配块．通过施加视角约束和顺序性约束，即平行结构摄像机产生的成像点的位置范围和顺序关系，对高斯先验概率计算模型进行修正，对不符合这二个约束的候选块惩罚．约束减少了扫描范围，从而减少了时间复杂度和产生误匹配的概率．实验结果表明，新方法的PSNR比相关法有02~4.0dB的提升，特别对噪声高、视差大的立体图像对提升较大．另外，视差图的光滑性要比相关法好．

New algorithm for stereo disparity estimation

To reduce the false match rate so as to increase Psnr and reduce computing complexity,a new algorithm for disparity estimation of the stereo image pair is proposed.Continuity and uniqueness constraints are employed to build up the Gaussian model for the computation of prior possibility,and the best matching block is determined according to maximum a posteriori(MAP) criterion.The angle of view constraint proposed in this paper and the ordering constraint,which are the pixels' position range and sequence relation,are imposed to correct the Gaussian prior possibility calculation model which gives punishment to the candidate blocks that violate these two constraints.Constraints reduce the scanning range so that computing complexity and possibility of false matching are both reduced.Experimental results show that the proposed algorithm gives PSNR about 0.2~4.0 dB upgrade in contrast to the normalized correlation.In particular,images with more noise or bigger disparities will be given more upgrade.In addition,the disparity map obtained from the proposed algorithm has better smoothness than that from the normalized correlation.