Controllable Probability-limited and Learning-based Human-like Vehicle Behavior and Trajectory Generation for Autonomous Driving Testing in Highway Scenario
Cheng Wei,Fei Hui,Asad J. Khattak,Yutan Zhang,Wenbo Wang
DOI: https://doi.org/10.1016/j.eswa.2023.120336
IF: 8.5
2023-05-06
Expert Systems with Applications
Abstract:Virtual simulation testing (VST) has become the main testing method for autonomous driving systems (ADSs) and autonomous driving assistance algorithms (ADAAs). The behavior and trajectory (B&T) information of a background vehicle in VST can provide different test scenarios for ADSs and ADAAs. Also, the B&T data on a differentiated vehicle can expand the test scope, increasing the reliability of the tested ADSs and ADAAs. Since human drivers have different driving styles, human-driven vehicles are one of the best choices for background vehicles. However, the VST based on natural driving data injection faces the problems of a small amount of data and difficult extraction of key samples and scenarios. Moreover, some of the current B&T generation methods are mostly targeted at autonomous vehicles, which are difficult to transfer to background vehicles. To address the mentioned shortcomings, this study proposes a method for generating human-like B&T data in batch. First, a uniformized velocity and trajectory sampling method with equal-length high-fidelity properties is developed for unequal-length natural driving data. Next, the lane-level vehicle entrance velocity is obtained, and a two-level probability-limited vehicle behavior generation and filtering method is developed. The generated behaviors are used as input to the trajectory generation model. Then, an enhanced learning-based trajectory generation model is designed, and trajectory generation is performed for different motion states using the generated behavior data. Finally, the coverage analysis of the proposed B&T generation model is performed, and the proposed model is compared with state-of-the-art B&T generation models. The results show that the proposed B&T generation model can achieve lateral and longitudinal TFRs of 87.85% and 88.45%, performing better than the other models. In addition, the proposed method can generate B&T data that are not included in the original data, solving the problem of the B&T data gaps and thus improving the B&T coverage of the VST background vehicles.
computer science, artificial intelligence,engineering, electrical & electronic,operations research & management science