high performance computing on graphics processing units: hgpu.org

Stereo matching is one of the first steps in the process of calculating 3D information from two 2D images. To triangulate a 3D point from two corresponding 2D features, the displacement in pixels, or the so-called disparity, must be estimated. From the estimated per-pixel disparity, using a projective camera model, 3D data for large portions of an image may be calculated. The 3D scene information can be used in applications ranging from obstacle detection and collision avoidance systems in the automotive industry to pick-and-place or human safety systems in the robotics industry. As time is an important factor in most of these applications, the subject of real-time stereo matching has gained importance while quality and accuracy aspects retain their importance. Benchmarks such as the KITTI Benchmark or the Middlebury Benchmark aim at providing stereo test data as well as ground truth to evaluate different matching algorithms against each other with regard to accuracy, coverage and runtime. However, they fall short in measuring the computational efficiency as the reported runtime as well as the real-time capability of the listed stereo matching algorithms are highly hardware dependent. In this thesis we explore the possibilities of real-time stereo matching and the constraints imposed by the used hardware. Therefore, we implemented a stereo matching algorithm in Open Computation Language (OpenCL) in order to evaluate the runtime of a specific algorithm on multiple devices. Using this runtime data, we discuss the limitations of runtime measurements with respect to varying computational power. Further, we suggest a method to compare the efficiency of various algorithms based on the reported runtime and hardware data, which is provided by the Middlebury Benchmark. This enables us to estimate the real-time capability of a given algorithm with a known problem space size on an arbitrary device with a manufacturer specified or measured performance figure. Finally, we observe that the problem space size, the device performance figure and the algorithm´s runtime complexity directly correlate with the matching rate given in Frames per Second (FPS).