Research On Dense Stereo Matching Between Aerial Epipolar Images Based On Multi-Channel Belief Propagation Algorithm

With the rapid development of society, there are urgent needs for real spatial3-dimentional information to sustain practical applications, such as life services? science research, and decision support. Of all types of spatial information carriers, optical images are easiest-requiring only basic surveying and mapping data. Aerial images are plentiful and widely used given the rapid development of aerial technologies. Problems which need to be urgent solved, however, include the effective use aerial images and techniques to convert image data into real spatial information that provides basic data support for science research.Traditional Photogrammetry research typically focuses on feature point detection technologies to search corresponding pixels. However, these detected corresponding pixels are always few in number, and therefore, utilization of images is low. Separately at the same time, many dense stereo matching technologies using dense matching between rectified pictures were developed in the Computer Version (CV) field. These technologies can identify large quantities of corresponding pixels from rectified pictures, allowing greater use of images. Thus, there is an unexplored potential to apply dense stereo matching technologies to indentify a large number of corresponding pixels in aerial epipolar images, this method provides a new solution for spatial information automatic acquiring from aerial images.Dense stereo matching technologies are often applied in low-resolution picture processing in CV, and are rarely applied in high-resolution aerial image matching processes. Dense stereo matching between aerial epipolar images is somewhat different from dense stereo matching in CV field. The differences, are not only caused by the adaptation of algorithms, but also caused by aerial images themselves. There is more interference in aerial images, given that:the resolutions are much higher, the corresponding pixels have serious color deviation, larger texture-less areas or texture-repetitive areas, moved objects, and so on. Taking the matching accuracy, expansibility, and computational acceleration feasibility of the dense matching algorithm into consideration, this paper presents a belief propagation algorithm for basic theoretical research and experimental verification. The research work in this paper has the three following parts.First, a multi-channel belief propagation theory is developed for dense stereo matching between aerial epipolar images. Evolved from the traditional Belief Propagation (BP) algorithm, a multi-channel belief propagation algorithm makes full use of all the three color components, R, G and B, and simplifies the process. The impact of the obvious color differences between corresponding pixels is the greatest interference in aerial image matching processing. In order to reduce this impact, a multi-channel belief propagation algorithm reduces the sensitivity of central pixels and increases the contribution of neighboring pixels.Second, many control technologies are integrated into the proposed multi-channel belief propagation algorithm for accuracy improvement. Based on the basic multi-channel belief propagation algorithm, a color segment result is used for controlling object boundaries, a window-based matching cost generation technology is applied to suppress interference caused by serious color deviation, and a disparity space images post-processing technology is applied for disparity classification and sub-pixel estimation.Three, parallel computational technologies are tested. The execution time of BP based algorithms is always very long and the resolution of aerial images is very high, this means that a multi-channel belief propagation algorithm needs computation acceleration technology to reduce execution time. Task parallelism and data parallelism are tested and compared. Both technologies can reduce execution time drastically. Data parallelism provided a larger speedup. Because of image data’s distributed storage and distributed processing, data parallelism is more suitable for current large parallel hardware, especially for multi-node-containing computers.