Ab­stract

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.