Texture Region based Hybrid Stereo Matching
Dongyeong Choi, Seong-Won Lee · IEIE Transactions on Smart Processing and Computing · 2018
Accurate disparity estimation is crucial for extracting depth information from stereo vision. However, in low-textured regions where homogeneity of pixel values is strong, the accuracy of disparity estimation is poor. Local disparity estimation methods have low accuracy on the lowtextured regions,. Global methods are more accurate for estimating disparity than local methods. However, they are too time-consuming, making them difficult to use. In this paper, we propose a low-textured region detection method based on energy in a specific frequency range. We use hierarchical frequency analysis of the pixel value variation within a window to improve disparity estimation. We also propose a hybrid stereo-matching method using feature descriptors in detected low-textured regions based on semi-global matching. It shows good disparity estimation accuracy. The scale-invariant feature transform feature descriptor has the best performance in our lowtextured region disparity estimation method, showing over 80% accuracy in low-textured region detection. Its disparity estimation performance improved in proportion to the ratio of low-textured regions, except for some images with large-sized, continuous low-textured regions.