The scenario-based tests have been an important means of improving the safety and reliability of autonomous vehicles (AVs). Moreover, ensuring the fidelity of generated scenarios is a safeguard for scenario-based tests. This paper presents a new approach to generate twin scenarios for the virtual testing of AVs using meta scenarios. By extracting key parameter clusters from real-world scenarios and redefining and reassembling them, a large number of twin scenarios are generated. This approach addresses the need for diverse scenarios in AV virtual testing and enhances efficiency and hazard detection in the Safety of the Intended Functionality (SOTIF) assessment. The validity of twin scenarios, their temporal logic, and social attributes are verified using the UPPAAL tool, ensuring reliability and fidelity in virtual testing with AV digital twins. This innovative method offers a valuable solution to address challenges in AV virtual testing and improve the overall assessment process.
Safety of the intended functionality (SOTIF) is paramount in autonomous driving, particularly at and beyond Level 3. An essential aspect of SOTIF research involves constructing and accurately assessing a broad range of hazardous scenarios. The current research landscape presents challenges in extensively testing autonomous vehicles and identifying all the unknown triggering events within their intended functionalities. Filling this gap, this article introduces a new methodology, noisy-HBN, driven by four-layer factors aggregated through the Internet of Vehicles (IoV), which integrates the Bayesian network with noisy nodes to assess the impact of unknown causes on SOTIF. To appraise the efficacy of this model, a new virtual data set, DADA-Plus, was developed, an extension of the open-source DADA data set, created using the twin scenario generation algorithm. The scenario-based verification underscores that the noisy-HBN model is adept at estimating the influence of unknown factors. Furthermore, the sensitivity analysis reveals that the known factors identified significantly correlate with the occurrence of accidents. Comparative analysis results show that the risk assessment coverage of the noisy-HBN has reached the state-of-the-art (SOTA) level.
Safety of the intended functionality (SOTIF) is expected to be the third type of road vehicle auto-driving safety in addition to information security and functional safety. SOTIF which is unknown and uncertain is caused by non-system failure or personnel misoperation. Advanced driver assistance systems (ADAS) that SOTIF cannot be accepted will be dangerous. The unknown triggering events triggered suddenly or exploited by a malicious hacker will pose a threat to vehicle safety. In the field of traditional network security, there have been many pieces of research on vulnerability assessment. This paper establishes hazard graph for SOTIF and triggering event models and proposes a safety assessment method for unknown triggering events of SOTIF and a method to enhances SOTIF inspired by the zero-day vulnerability assessment in the traditional cybersecurity field. The AEB system is used as an example to apply the proposed algorithm, and Petri net accessibility testing is used as a verification of the correctness of the algorithm in this paper.
Detecting and understanding driver behaviors in automated vehicles are critical for reducing the risk of misuse and ensuring the Safety Of The Intended Functionality (SOTIF). Using a Temporal Recurrent Network (TRN), the paper proposes the driver-TRN model to detect and predict driver behaviors based on real-time vehicular data, such as in-car video, road camera, and sensor data, which addresses intended risks resulting from the interaction between drivers and automated systems that may be overlooked. This paper improves the performance of the driver-TRN with multiple data sources and multi-task loss fine-tuning. The experiments based on the Brain4Cars dataset show that the driver-TRN can enhance the safety and reliability of automated vehicles by providing a more comprehensive detection of driver behaviors.
Safety of the intended functionality is an indispensable safe condition that must be considered for autonomous driving on the road. The key to study SOTIF is to construct as many hazardous scenarios as possible and accurately assess SOTIF. It is difficult to test many automated driving vehicles in the current research and discover all the unknown triggering events in the intended functionality. As a new simulation method, vehicle digital twins can simulate as many driving scenarios as possible efficiently and at a low cost in a short time. Therefore, this paper proposes an algorithm to assess SOTIF with unknown hazards based on the vehicle digital twins based on a Bayesian hazard graph with attributes and Noisy-OR structures. This paper uses CARLA to build the vehicle digital twins and testing this algorithm based on these digital twins. After testing, the SOTIF quantitative assessment algorithm proposed in this paper is effective and has good performance.