**6. Conclusions**

In order to support efficient virtual test and validation of LKA systems, this study developed an MLP model to determine lane detection *C*0-LPE and *C*1-HAE estimation based on the relationship of vehicle dynamic data. This relationship is complex for an actual dataset derived from real-world measurements and requires an artificial intelligence method to create a reliable model to analyze the problem. This approach was divided into three parts. Firstly, the measurements and data collection were carried out for the testing procedure, and digital twin-based data were defined as ground truth. The second part was to extract data and select features from the actual collected data to find input data that had greater influences on the model, thus improving training efficiency. In the third part of the study, an MLP model was developed, and the selected features were used as inputs to train the model. The results also showed that MLP can produce higher accuracy than other regression approaches. Finally, the technique was employed to reproduce lane detection behaviour of an automotive camera system in a simulation platform. Combined with the analysis of the simulation results, we found that the best regression is achieved for a given non-linear dataset. Due to the fact that existing data and tests were conducted primarily on straight roads, lane marking detection on curved roads will be taken into account to refine the model further and improve our approach.

The model fits the detection error of the sensor output by using selected features, which enables fast and efficient sensor modelling. Compared to the physical model, this approach simplifies the modelling process by ignoring physical performance modelling of the camera components as well as the perception algorithm and focusing only on the inputs and outputs of the camera system, thus improving computational performance. Moreover, in contrast to the ideal models previously mentioned, ideal sensor models provide only ground truth information without any specific post-processing function. Therefore, physical effects do not influence these models. However, PLDM models based on the MLP approach can provide more details about sensor detection performance than an ideal model, enhancing the simulation's realism. Although there is a strong correlation between modelling complexity, training time, data composition and volume, modelling efficiency is improved, and this approach is generic. It can be applied to various sensors with low efforts after initial development.

**Author Contributions:** Conceptualization, H.L.; methodology, H.L. and S.A.; software, H.L. and K.T.; validation, H.L. and K.T.; investigation, H.L.; resources, D.B. (Darko Babic), D.B. (Dario Babic), Z.F.M. and V.T.; data curation, H.L., K.T. and C.W.; writing—original draft preparation, H.L. and K.T.; writing—review and editing, H.L., S.A. and A.E.; visualization, H.L.; supervision, A.E. and M.C.B. All authors have read and agreed to the published version of the manuscript.

**Funding:** Open Access Funding by the Graz University of Technology.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** The author thanks those who have supported this research, to the Graz University of Technology also.

**Conflicts of Interest:** The authors declare no conflict of interest.
