Multi-Platform Performance Evaluation of Pedestrian Detection at the Electronic System Level

Abstract:
The growing complexity of hardware platforms for driver assistance systems makes a cost-efficient selection and algorithm implementation challenging. Here, Electronic System Level tools, which combine model-based de-sign, system synthesis, and performance evaluation, seem to be a remedy. In this paper, we systematically assess the accuracy of system level performance evaluation through a quantitative comparison of estimated and meas-ured performance numbers. As a case study, we use a pedestrian detection application. The application is mapped onto three different reference platforms, an FPGA, an ASIP, and a multi-core processor. The average error of the system-level performance estimation was typically between 3% and 16% at a standard deviation of below 11%.