(A) real-time stereo depth estimator based on image pyramid

Abstract

Stereo matching is used in many real-time applications that require fast and relatively accurate matching algorithms and simple hardware architectures. Although the semi-global matching algorithm performs better than the window-based local matching methods with the same disparity search range, it still requires a significant amount of computation and storage spaces, not satisfying real-time performances. In addition, although the disparity search range represents the level of depth estimation, it is limited to small values in semi-global matching due to these overheads, which can not accurately estimate near objects with large disparity values. This paper introduces the image pyramid method into stereo depth estimation to reduce the computation and storage requirements required for stereo matching while maintaining a wide search range. Also, this paper addresses the issue of disparity errors caused by introducing different image sizes, minimizing performance degradation. The proposed image pyramid-based matching algorithm employing census transform and four paths semi-global matching achieves an error rate of 27.3% in non-occlusion regions on the Middlebury 2014 dataset while reducing the computation required to calculate and aggregate the matching cost for searching disparity values to 24%. The depth estimator synthesized in 65 nm technology for that algorithm attains 87 frames per second on Full HD images with a 512 disparity search range.