Generative AI for Lateral Movement Modeling in Comparison to Classical Modeling Approaches

The lateral offset between vehicles is a decisive input for sensor range of vision estimation and thus a relevant parameter for the maximum situational awareness an automated driving function relying on such sensors can achieve. With simulations playing an integral role in the validation process of automated driving functions, the ability to realistically model the lateral movement of vehicles within their lane is crucial. Previous studies therefore proposed models for lateral movement modeling based on stochastic differential equations and twolevel approaches combining a Markov model and a noise model. Within this work, we present a deep state space model for generating lateral movement profiles and compare it with the previously mentioned classical approaches. The deep state space model outperforms the models based on stochastic differential equations. While it does not reach the performance of the two-level model in every regard and also does not fulfill all requirements on lateral movement modeling identified, it achieves a high accuracy with strengths different from the two-level model. It is therefore a valuable enhancement for lateral movement modeling and can serve as a reference for evaluating classical approaches.