基于递推自适应权重的快速稠密立体匹配

针对经典自适应权重稠密立体匹配算法计算量大的问题,提出了一种递推自适应权重算法.重新定义相邻像素的权重为距离衰减因子和色彩差异函数的乘积,不相邻像素权重为相邻像素权重的累乘,色彩差异越小、距离越近的像素权重越大;证明了在新的权重定义下,一维空间的匹配代价融合可以通过两次递推完成,真实图像的匹配代价融合可以通过4次递推完成,同时给出相应递推公式;递推匹配代价融合时每个像素每一视差只做4次乘法和8次加法,计算量比窗口大小为35×35的经典自适应权重算法小约两个数量级;基于递推匹配代价融合实现了一种快速稠密立体匹配算法.使用Middlebury大学的测评集测试该算法,证明了递推自适应权重算法的速度和精度均优于经典自适应权重算法.

Fast recursive adaptive weight stereo matching

Stereo matching based on traditional adaptive weight is computational intensive. The basic idea of adaptive weight is that bigger weight should be given to those pixels with less color difference and shorter distance. A novel weight was defined to recursively implement cost aggregation. The weight between neighbor pixels was redefined as the product of distance attenuation factor and color difference function, while the weight between other pixels was redefined as the product of weights between neighbor pixels. Using the proposed weight, cost aggregation was recursively implemented with only 4 multiplications and 8 additions per pixel per disparity. A new fast dense stereo matching was designed based on recursive adaptive weight. Evaluation on the Middlebury’s benchmark proved that the proposed method is faster and more accurate than traditional adaptive weight method.