为了准确获取实际场景中的深度信息,本文通过引入空间距离权重,提出了同时考虑局部区域相似度和接近度的加权支持域立体匹配方法。首先,对输入图像进行滤波,去除图像中的噪声,并通过Mini-Census变换求取Hamming距离。然后,建立加权支持域,求取代价累积;进而通过winner-take-all方法求取最小匹配代价和原始视差图。最后,对原始视差图进行细化处理,得到优化的视差图,并反演出空间的深度分布。实验结果表明,利用该算法在不同光照、不同场景下均能够正确地产生视差图;对标准数据库图片进行处理时,平均错误率仅为6.77％。该算法有效地降低了计算复杂度,具有计算精度高、适应性强、鲁棒性好、便于硬件实现等特点, 可为高精度实时立体匹配专用处理硬件的设计和实现提供基础。

A hardware friendly stereo matching algorithm by employing a weighted support region was proposed with considering the local similarity and proximity of the local region to acquire the depth information in the real scenes accurately. Firstly, the input image were filtered to remove the noise and the results of the Mini-Census transform were used to calculate the Hamming distance. Then, the weighted support region was built to finish the cost aggregation and furthermore to find the minimal matching cost and generate a raw disparity map by the winner-take-all method. Finally, the raw disparity map was refined to generate the optimal disparity map and to deduce its depth distribution in the real scenes. The experimental result shows that the algorithm produces the disparity map accurately in different illumination conditions and different scenes, and its processing average bad pixel rate is only 6.77% when the images of the standard database are processed by the proposed algorithm. The proposed algorithm reduces computational complexity and is characterized by higher accuracy, good adaptability and fine robustness. It provides a basis for design and implementation of specific hardware for high accuracy stereo matching.