Comparison of FPGA and GPU implementations of real-time stereo vision

Real-time stereo vision systems have many applications - from autonomous navigation for vehicles through surveillance to materials handling. Accurate scene interpretation depends on an ability to process high resolution images in real-time, but, although the calculations for stereo matching are basically simple, a practical system needs to evaluate at least 109 disparities every second - beyond the capability of a single processor. Stereo correspondence algorithms have high degrees of inherent parallelism and are thus good candidates for parallel implementations. In this paper, we compare the performance obtainable with an FPGA and a GPU to understand the trade-off between the flexibility but relatively low speed of an FPGA and the high speed and fixed architecture of the GPU. Our comparison highlights the relative strengths and limitations of the two systems. Our experiments show that, for a range of image sizes, the GPU manages 2 × 10⁹ disparities per second, compared with 2.6 × 10⁹ disparities per second for an FPGA.