Performance characterization of automotive computer vision systems using Graphics Processing Units (GPUs)

The purpose of this paper is to propose a method to abstract and classify vehicle data collected from vision sensors into road scenarios. The classified scenarios can be played back on specialized hardware designed to handle these scenarios to characterize its performance. Since the majority of existing automotive computer vision systems mandate real-time results, this study aims to introduce the utilization of Graphics Processing Units (GPUs) as a prototype to perform these classification and abstraction tasks. This paper evaluates the ability of the GPU architecture to handle these tasks. It also discusses the suitability of GPUs for integrating navigation data with data from vision and RADAR sensors for aiding Visual Simultaneous Localization and Mapping (V-SLAM) tasks for future autonomous vehicle platforms.