Review on Functional Testing Scenario Library Generation for Connected and Automated Vehicles

Abstract

The advancement of autonomous driving technology has had a significant impact on both transportation networks and people’s lives. Connected and automated vehicles as well as the surrounding driving environment are increasingly exchanging information. The traditional open road test or closed field test, which has large costs, lengthy durations, and few diverse test scenarios, cannot satisfy the autonomous driving system’s need for reliable and safe testing. Functional testing is the emphasis of the test since features such as frontal collision and traffic sign warning influence driving safety. As a result, simulation testing will undoubtedly emerge as a new technique for unmanned vehicle testing. A crucial aspect of simulation testing is the creation of test scenarios. With an emphasis on the map generating method and the dynamic scenario production method in the test scenarios, this article explains many scenarios and scenario construction techniques utilized in the process of self-driving car testing. A thorough analysis of the state of relevant research is conducted, and approaches for creating common scenarios as well as brand-new methods based on machine learning are emphasized.

Categorizing Data-Driven Methods for Test Scenario Generation to Assess Automated Driving Systems

“…In [130], dynamic scenario generation is divided into four methods: combinatorial testing, knowledge-and drivingbehavior-based generation, and data-driven scenario generation. Roads were created independently from dynamic scenario content using field-collection data from OpenStreetMap files or remote sensing imagery.…”

Augmenting Scenario Description Languages for Intelligence Testing of Automated Driving Systems

“…However, this is insufficient as, e.g., CF 1 could be true if the vehicle V1 is detected only for a single time frame. Hence, we define complementary constraints, i.e., CF (11)(12)(13)(14), to "ensure" continuous success of the intelligence functions from the moment they become active to the appropriate end boundaries, e.g., end of the scenario. For example, CF 1, CF 2, CF 11 and CF 12 together "ensure" the vehicle V1 is detected and recognized correctly throughout the scenario.…”

“…Compared to the distance-based approach where the ADSs are required to drive millions of miles [4], [5] to cover sufficient diversity of driving situations due to the long-tail effect, scenario-based methods aim to eliminate the redundancy and distil critical scenarios of interest directly from various data sources, e.g., domain expert knowledge [6]- [12] or naturalist driving data [12]- [24] by diverse types of scenario generation (aka, parameter sampling) algorithms. The generated scenarios are commonly evaluated against two metrics, i.e., criticality (e.g., distance-to-collision [22], [23] and time-to-collision [10], [12], [21], [25]), and coverage (e.g., parameter value combination [11], [12]).…”

Optimal method of extreme scenarios for intelligent driving vehicle testing based on time window

“…These include two main methods for constructing different types of scenarios. One of them is replicating frequently occurring situations encountered during real-life driving experiences [1] . The other kind of methods is generating rare but highly critical hazardous scenarios specifically [2,3,4] .…”