Simulation-based Methods for the Validation of Highly Automated Driving

In recent years, the automotive industry has striven to introduce automated driving functions (ADFs) of increasing complexity and capability. Traditional statistical validation methods like endurance runs, which were previously applied to prove automotive safety, reach their limits and become infeasible since society poses increasingly challenging safety demands to increasingly complex and capable ADFs. Simulations promise to deliver a valuable contribution to the resulting unsolved validation challenge, since they show beneficial properties like being reproducable, safe, controllable and fast. This thesis addresses some of the most pressing issues hindering the usage of simulations for validation purposes. Firstly, a sufficiently well performing ADF fails seldomly. Hence, critical scenarios occur very rarely in statistical simulations and a high amount of simulation runs is needed to generate results with acceptable accuracy. Importance sampling (IS) was previously used to tackle this problem. IS skews the sampling of simulation runs to the areas of the parameter space which the ADF cannot handle, but allows determining unskewed statistical results. A typical approach to identify such critical areas in the context of IS is the cross-entropy method (CE). CE’s optimization procedure scales badly with the complexity of modeled scenarios. This thesis introduces a new framework which allows identifying and limiting the CE optimization on relevant parts of parameter space. It is shown, that this enables scalable application of IS to complex scenarios as found on public roads. Secondly, simulations require statistical models which represent the scenes in the surrounding of the tested ADF. Previously available scene models concentrated on simple road topologies like straight highways sections. Within this thesis, existing models are generalized to more complex topologies as intersections. Initially, the proposed model is based on Bayesian networks (BN). However, BNs scale badly with the number of learned parameters. Thus, the model is adapted to the usage of sum-product networks (SPNs). The proposed scene model is statistically evaluated and it is shown that SPNs increase the model’s scalability. Thirdly, it is shown how the proposed scene model can be combined with IS. The scene model’s structure is adapted to incorporate parameters which address a scene’s criticality. As it is proven, that enables CE to skew the model’s inherent statistics in order to sample predominantly critical scenes and to quickly generate statistical evidence about an ADF’s safety.