Ab­stract

Com­pre­hen­sive per­cep­tion of the ve­hi­cle’s en­vi­ron­ment and cor­rect in­ter­pre­ta­tion of the en­vi­ron­ment are cru­cial for the safe op­er­a­tion of au­tonomous ve­hi­cles. The per­cep­tion of sur­round­ing ob­jects is the main com­po­nent for fur­ther tasks such as tra­jec­tory plan­ning. How­ever, safe tra­jec­tory plan­ning re­quires not only ob­ject de­tec­tion, but also the de­tec­tion of dri­vable areas and lane cor­ri­dors. While first ap­proaches con­sider an ad­vanced safety eval­u­a­tion of ob­ject de­tec­tion, the eval­u­a­tion of lane de­tec­tion still lacks suf­fi­cient safety met­rics. For as­sess­ing safety, ad­di­tional fac­tors such as the se­man­tics of the scene with road type and road width, the de­tec­tion range as well as the po­ten­tial causes of miss­ing de­tec­tions, in­cor­po­rated by ve­hi­cle speed, should be con­sid­ered for the eval­u­a­tion of lane de­tec­tion. There­fore, we pro­pose the Lane Safety Met­ric (LSM), which takes these fac­tors into ac­count in order to eval­u­ate the safety of lane de­tec­tion sys­tems by de­ter­min­ing an eas­ily in­ter­pretable safety score. We eval­u­ate our of­fline safety met­ric on var­i­ous vir­tual sce­nar­ios using dif­fer­ent lane de­tec­tion ap­proaches and com­pare it with state-of-the-art per­for­mance met­rics.