| Binocular stereo,in recent years made considerable development, has been widely used in intelligent vehicle and other fields. As the rapid development of binocular stereo vision system, the precision standard gets higher and higher. This make the data processing also growing. The timeliness has greatly limited the technology use in the real world. The obstacle detection based on binocular stereo vision is a critical part of the technology. Based on the study of the stereo matching, we gave a new algorithm, and use the parallel computing to meet the real-time implementation.We first introduced the existing binocular vision’s theory and algorithms. And then we proposed a new-designed semi-match algorithm for the obstacle detection. We gave the obstacle detection implementation structure, including a Gaussian filter, gradient transform module, census cost of computing, SGM semi-global matching and edge detection module. As the algorithm need a large number of filtering and caching, a high degree of complexity requires a lot of cache data and calculations, we proposed a hardware design using the paralle computing for this algorithm. In the census cost computing module, we changed the objects of the comparison to make the cost represent the match pixel’s neighborhood information more effectively. The traditional matching algorithm acceleration in hardware can only take advantage of the pixel’s neighborhood gray value, this paper use a plurality of one-dimensional direction dynamic programming, by the Semi-global matching(SGM) module, to increase the of of inconsistent light conditions and the other noise.The RTL code of the whole module was described by Verilog HDL language. We used the matalb for the algorithm’s function verification, and the ISE to simulation the RTL module. Finally, we synthesized the RTL code of the obstacle detection by ISE. The highest frequency of the design was up to 28.817 MHz, fully meet the requirements of real-time image processing. |
