基于CUDA加速的自适应窗口匹配算法

Stereo Matching Algorithm by CUDA-Accelerated Adaptive Window

通过双目立体视觉计算视差与深度是三维重建、虚实交互等多种应用的重要基础,基于窗口的自适应匹配算法准确度高,但计算复杂,难以应用于实时系统中。为提高算法效率,提出基于CUDA加速的自适应窗口匹配算法,根据颜色相关性与几何近似性计算中心像素点之间的匹配程度,利用GPU的流多处理器和共享显存,将搜索窗口匹配核函数分配到多个并行执行的线程中,各个像素点的视差值并行计算,减少计算时间。对国际标准数据集图像的实验结果表明该算法在保证较好的匹配精度的同时,计算效率显著提高。

Disparity and depth computation through binocular vision is very important in multiple applications including three dimensional reconstruction and virtual-reality interaction. Window-based adaptive correspondence algorithms have high accuracy as also as high computation complexity, so they can hardly be applied to real-time systems. In order to improve the efficiency, a stereo matching algorithm was proposed by＂ CUDA-accelerated adaptive window. It made use of color correspondence and geometric similarity to calculate the matching degree between central pixels. Using stream multiple processors and shared memory of GPU, the kernel functions of window correspondence were allocated to multiple threads which were conducted in parallel and the disparities of each pixel were calculated independently to reduce the computation time. Experiments on the international standard image sets show that the algorithm has greatly enhanced the efficiency while keeping good matching precision.