Autonomous driving hazard scenario extraction and safety assessment based on crash reports and carla simulation

Abstract

To solve the traffic safety problems that arise during "mixed traffic" between autonomous and manual vehicles, this paper presents a harzard scenario simulation and safety assessment method based on autonomous driving crash reports. 197 crash reports of autonomous vehicles are selected from California Department of Motor Vehicles and used for autonomous driving hazard scenario extraction. Combining 37 pre-collision scenarios published by the National Highway Traffic Safety Administration, typical traffic safety scenarios are identified. A random forest model is used to identify key elements of typical autonomous vehicle accident scenarios. The CARLA microscopic traffic simulation platform is used to simulate two typical "mixed traffic" safety scenarios: Lead Vehicle Stopped and Following Vehicle Making a Maneuver. Three indicators -relative velocity, relative distance, and time-to-collision -are selected for constructing a safety evaluation suitable for autonomous vehicles. The research results show that in the Lead Vehicle Stopped scenario, when the following vehicle is 15 meters away from the front vehicle, the risk of collision is extremely high when the initial speed of the following vehicle is greater than 13.1 m/s. In the Following Vehicle Making a Maneuver scenario, when the initial speed of the following vehicle is greater than 12.5 m/s, the risk of collision with the front vehicle is also extremely high. The results can provide support for the safety evaluation of autonomous driving and the improvement of autonomous driving system.