Performance analysis of multi-resolution symmetric dynamic programming stereo on GPU

High precision 3D scene maps are essential for many computer vision based applications such as e.g. autonomous vehicle navigation and collision avoidance. However, stereo vision is an ill posed inverse optical problem with only a few viable regularised solutions making generation of depth maps with large disparity ranges (128 - 512 pixels) from high resolution (mega-pixel) stereo images too computationally intensive. Although hierarchical stereo matching can help in efficient handling of large disparity ranges, the effective use of parallel hardware to provide high frame rates still is challenging. We describe and analyze a hierarchical implementation of the symmetric dynamic programming stereo (SDPS) algorithm on a cheap commercial GPU. Our approach not only reduces the overall computation demands and thus increases achievable frame rates, but also it uses inter scan-line consistency to improve matching performance and incorporates scene constraints directly into the disparity calculation. Experiments with a GTX 295 GPU for different image and disparity resolution combinations gave promising results: e.g. 4-Megapixel images with the disparity range of 256 can be processed at 14 frames per second. The performance for different configurations is analyzed and compared to other (FPGA and CPU) implementations.