特征融合的双目立体匹配算法加速研究与实现

Research and Implementation Accelerating of Binocular Stereo Matching Algorithm Based on Feature Fusion

关键词:  立体匹配  SAD  Census变换  CUDA加速  并行计算

Key words:  Stereo matching  SAD  Census transform  CUDA acceleration  Parallel computing

摘要:

随着图像分辨率和场景信息获取实时性需求的提高，业界对双目立体匹配算法的效率提出了更高的要求。针对该问题，提出了将SAD与Census变换特征融合的结果作为初始匹配代价，利用SGM算法进行代价聚合，采用赢家通吃策略计算视差，通过左右一致性检验检测出遮挡点并填充，使用中值滤波剔除异常值，最终获取优化后的视差图。采用统一计算设备架构（CUDA）对算法实现并行计算，针对立体匹配比较耗时的问题，该算法最大化地利用共享内存、寄存器内存以及CUDA流，实现了不同核函数之间的并行，大大提升了执行效率。结果表明，该算法在Middlebury立体匹配平台上，平均误匹配率下降了8.05%；在NVIDIA GeForce GTX 1650平台上运行450×375分辨率的图像，比原始SGM算法快687倍,运行高分辨率图像时依然能够实现实时显示性能。

Abstract:

With the increase in image resolution and the demand for real-time scene information acquisition, the industry has put forward higher demands on the efficiency of binocular stereo matching algorithms. To address this problem, the outcome of the fusion of SAD features and Census transform is proposed as the initial matching cost, cost aggregation is performed using SGM algorithm, parallax is calculated using a winner-take-all strategy, occlusion points are detected and filled by left-right consistency check, outliers are removed using median filtering, and finally the optimized parallax map is obtained. The CUDA is used to parallelize the algorithm. For the problem that stereo matching is relatively time-consuming, the algorithm maximizes the use of shared memory, register memory and the use of CUDA streams to achieve parallelism among different core functions, which greatly improves the execution efficiency. The results show that the algorithm reduces the average false match by 8.05% on the Middlebury stereo matching platform; it is 687 times faster than the original SGM algorithm when running 450×375 resolution images on the NVIDIA GeForce GTX 1650 platform， and it still achieves real-time display performance when running high-resolution images.