With the advent of semi-autonomous vehicles on the roads, questions about their reliability arise. Due to the significantly improved environmental perception, the complexity an algorithm needs to deal with has drastically increased over the last years. Utilizing sensors such as lidar, radar, and camera, the number of system states a safety critical algorithm has to cover became almost infeasible for validation on the test track.
In this paper, a simulation framework is presented, which is dedicated to identifying challenging situations for vehicle safety algorithms. The results can be used to concentrate the test-track efforts to scenarios, which are most likely to trigger an undesired system behavior. Situations of interest are those, which cause obvious misbehavior of the system-under-test, as well as unrobust situations resulting in large changes in the output when the input is only slightly varied.
The framework comprises multiple modules responsible for a continuous simulation of the pre-, in- and post-crash phases. The open source traffic simulation tool SUMO provides realistic traffic scenarios, while the combination of a vehicle dynamics model and a newly developed mass-spring model are used to approximate the vehicle movement even during the In- and Post-Crash phases. Driver actions, as well as sensor information, are varied in multiple instances of the framework in order to guarantee a high coverage of the possible system states. The scenarios are always motivated by a reasonable “real world” background, rather than random sampling.
With a focus on crash scenario analysis, a rule-based filter ensures that preferably situations with a high collision probability are chosen. Using this technique, millions of random crash situations are generated within days, where a comparable number would take years to observe in the real world.
The framework is used in the development process of pre-crash functions. It successfully is employed to evaluate the quality of predictive algorithms dealing with the uncertainties of future events. The data generation aspect of the framework is utilized for vehicle safety-related machine learning tasks.
«
With the advent of semi-autonomous vehicles on the roads, questions about their reliability arise. Due to the significantly improved environmental perception, the complexity an algorithm needs to deal with has drastically increased over the last years. Utilizing sensors such as lidar, radar, and camera, the number of system states a safety critical algorithm has to cover became almost infeasible for validation on the test track.
In this paper, a simulation framework is presented, which is ded...
»