Robustness Evaluation and Improvement for Vision-Based Advanced Driver Assistance Systems

In this paper we propose a novel method of robustness evaluation and improvement. The required amount of on-road records used in the design and validation of vision-based advanced driver assistance systems and fully automated driving vehicles is reduced by the use of fitness landscaping. This is realized by guided application of simulated environmental conditions to real video data. To achieve a high test coverage of advanced driver assistance systems many different environmental conditions have to be tested. However, it is by far too time-consuming to build test sets of all environmental combinations by recording real video data. Our approach facilitates the generation of comparable test sets by using largely reduced amounts of real on-road records and subsequent application of computer-generated environmental variations. We demonstrate this method using virtual prototypes of an automotive traffic sign recognition system and a lane detection system. The robustness of these systems is evaluated and improved in a second step.