Реализация алгоритма сопоставления стереоизображений

Abstract

Recently, with the development of technologies such as augmented reality and unmanned vehicles that receive information about the environment using video cameras, there is a need for fast and accurate depth assessment in images. This problem is relevant, and has led to in-depth research in this area. This article discusses the fundamental principles of local and global stereo image matching algorithms. The operation of the Semi-Global Block Matching algorithm is described in more detail. The essence of the algorithm is to perform line optimization in multiple directions and calculate the aggregated costs by summing the costs of reaching a pixel with an inequality in each direction. The number of directions affects the execution time of the algorithm, and while 16 directions usually provide good quality, a smaller number can be used to achieve faster execution. A typical implementation of the algorithm in 8 directions can calculate the cost in two passes: the forward pass accumulates the cost on the left, the top-left, top and top-right, and the reverse pass accumulates the cost on the right, bottom. right, bottom, and bottom left. Its implementation is also affected by transferring the calculation of the depth map to a graphics processor (GPU) to speed up processing. The results of the construction of the depth map, as well as the dependence of the time of the algorithm on the size of the input images are shown.