Calibration and validation of software performance models for pedestrian detection systems

In recent years, road vehicles have seen a tremendous increase on driver assistance systems like lane departure warning, traffic sign recognition, or pedestrian detection. The development of efficient and cost-effective electronic control units that meet the necessary real-time performance for these systems is a complex challenge. Often, Electronic System-Level design tackles the challenge by simulation-based performance evaluation, although, the quality of system-level performance simulation approaches is not yet evaluated in detail. In this paper, we present the calibration and validation of a system-level performance simulation model. For evaluation, an automotive pedestrian detection algorithm is studied. Especially the varying number of pedestrians has a significant impact to the system performance and makes the prediction of execution time difficult. As test cases we used typical sequences and corner cases recorded by an experimental car. Our evaluation results indicate that prediction of execution times with an average error of 3.1% and a maximum error of 7.9% can be achieved. Thereby, simulated and measured execution times of a software implementation are compared.