摘要： 利用可编程图形硬件GPU 实现了非参数局域变换双目立体视觉算法。该算法使用局部非参数统计的结果而不是像素灰度值作为匹配代价，相对于其它基于区域的立体匹配算法，具有物体边界区域处理稳定和适于硬件实现等优点。该文利用GPU 的最新特性实现了算法的全部运算都在GPU 上执行。由于GPU 的并行流水特性，算法在GPU 上的运算速度较在CPU 上得到提高。

关键词: 双目立体视觉；图形处理器；非参数统计；视差图

Abstract: This paper presents a binocular stereo vision algorithm implemented on GPU. Compared with the other area-based stereo algorithm, it takes the relative order of local intensity values rather than gray values of pixels as matching primitive. This can result in improved performance near object boundaries in disparity map and being well-suited for hardware implementation. By taking advantage of the advanced features of recent GPU,the implementation runs entirely on standard graphics hardware. Thanks to parallel pipeline in GPU, the run time on GPU is a lot shorter than on CPU.

Key words: Binocular stereo vision; Graphics processing unit (GPU); No-parametric statistics; Disparity map