Verification of Autonomous Vehicles: Scenario Generation based on Real World Accidents

Abstract

With the advent of technologies that allow for selfdriving and autonomous vehicles, it is of the utmost importance to ensure the safety of occupants, pedestrians, and other motorists. Investigating when, why, and how these vehicles fail is integral to preventing future recurrences of these kinds of incidents. By gathering these accidents, their associated data, and simulating them, it is possible to identify the major reasons for the accident as well as allowing for the testing and verification of autonomous vehicles (AVs), before they encounter the real world. This knowledge can then be further refined for the purposes of establishing a set of standards for AV safety and capabilities. This paper proposes a novel system for crash scenario generation which works based on actual AV crashes. In addition, there does not yet exist a cloud based database for housing AV crashes, locations, and descriptions. Furthermore, there also does not exist an automated system that creates simulated testing scenarios based on real AV crashes. Our proposed algorithm checks the AV crash location and designs virtual roads from map data about the crash location, retrieved from Open Street Maps. Then it is planned to use natural language processing (NLP) to generate the crash scenario based on the crash description. The virtual scenario, which includes real features of roads – such as curvature, number of lanes, and speed limits – will be used to test AVs. Another portion of the proposed system will allow for important features to be extracted automatically from these crashes, which will be used for improving the road designs capable of addressing AV’s needs.