자율주행자동차의 추돌 회피를 위한 교통사고분석 및 기계 학습 기반 위험 시나리오 생성 연구

Abstract

In this paper, the critical scenario generation method for the scenario-based approach is proposed to validate collision avoidance systems on autonomous vehicles. Along with three abstraction levels of scenarios for the safety of the intended functionality(SOTIF), as proposed by a PEGASUS project in Germany, critical scenarios based on fatal traffic accidents in Korea were analyzed statistically. Then, the collision scenario model, including all critical scenarios, is proposed to generate logical scenarios systematically. Since the high dimension of parameters in a logical scenario results in a combinatorial explosion of concrete scenarios, it is quite necessary to search for appropriate scenarios. Therefore, many safe scenarios were omitted by applying for a series of conditions based on time-to-collision and support vector machines. Finally, It is shown how scenarios can be generated to validate an automatic emergency braking system, and the critical scenarios are searched out via the proposed generation procedure.