Controllable probability-limited and learning-based human-like vehicle behavior and trajectory generation for autonomous driving testing in highway scenario

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.