Parallel matching-based estimation - a case study on three different hardware architectures

Many advanced driver assistance systems (ADAS) and autonomous vehicles require 3D information available from (stereo) camera systems. The corresponding task of estimating disparity or optical flow is computationally demanding, so meeting real-time update rates at high image resolutions has proven to be challenging. Modern parallel hardware seems suitable for this task only if its processing power can be efficiently accessed by parallel software implementations. In this paper we present a case study comparing different hardware platforms by two variants of block matching-based estimation. These platforms include two x86-compatible multicore systems, a graphics processing unit (GPU) and a 64-core embedded design. We introduce relevant features of each architecture and describe their effects on the applied algorithms, parallelization approaches and implementations. Target platforms are evaluated concerning computational performance, energy efficiency and programmer productivity.