*Motivation*

Increasing the level of driving automation leads to an increase in system complexity. Therefore, the number of test cases necessary to proof the functional safety increases exponentially [5]. The literature reports that hundreds of millions of accident-free driving kilometres are needed to prove that the system is better than human vehicle control in terms of vehicle safety, e.g., [6]. Therefore, testing and validation efforts are being shifted towards virtual validation as well as X-in-the-Loop methods [7]. An essential requirement for testing ADAS/AD systems in virtual space is the realistic modelling of the virtual environment required for the system under test.

Virtual testing is particularly suitable for testing safety-critical scenarios that are difficult, costly, unsafe and impossible to reproduce on test tracks or roads [8]. The development and testing of driving assistance and automated vehicle control systems is performed step by step, from simple object detection to highly sophisticated functions, in the phases defined in the V-Model presented in ISO 26262-2:2018 [9]. Accordingly, the virtual environment, as well as the architecture and capabilities of the sensor models used, will vary according to the development phases [10].

To accelerate the test execution, a possible, and in recent years, very relevant solution is provided by the use of simulation tools on a virtual basis [11,12]. In the case of perception–sensor models in early development phases, virtual simulation based on ideal or phenomenological sensor models has become established in the industry. These models can be used to test and validate the fundamental operating principles of control architectures.

Using advanced perception–sensor models enables the testing of machine-perception and sensor-fusion algorithms in a later stage, enabling thus a first parameter tuning on a complete vehicle level before the first prototype is built.

Since sensor models have a limited ability to represent reality, careful consideration must be given to whether the model is a satisfactory replacement for the real sensor for validating the safety of ADAS/AD functions. However, there is no accepted methodology available that objectively quantifies the quality of perception–sensor models. In the present paper, a novel approach to assess the performance of virtual perception model is described.

The rest of the paper is structured as follows: Section 2 reviews the state-of-the-art technology, Section 3 describes the used method including the digital twin of the driving environment, the vehicle and the assessment approach. Section 4 presents the results of the new performance evaluation method using the IPG RSI model compared to an automotive radar sensor already proven on the market. Section 5 discusses the results of the sensor performance assessment and describes the limitations detected during the research and gives an outlook on future improvements.
