Advanced driver assistance systems (ADAS), for example autonomous emergency brake (AEB) systems, aim to sense a vehicle’s environment, understand the current traffic situation and react appropriately in order to support the human driver. One cornerstone for reliable systems is an appropriate handling of uncertainties. Uncertainties arise for instance due to noisy sensor measurements or the unknown future evolution of a traffic situation. This work’s objective is to contribute to the understanding of these uncertainties.
To this end, parametric probability distributions are used to analytically model and propagate uncertainties at individual parts of ADAS signal processing chains. Previous works approach this mostly with numerical simulations. An inherent drawback is that only a concrete implementation of algorithms can be analysed. Analytical modelling on the other hand, as pursued in this thesis, allows drawing more generic conclusions. One can for example attempt to derive upper performance bounds which apply to any implementation of (sub-) optimal solutions.
This thesis devises probabilistic models of uncertainty in selected algorithms with a distinct role in current and future ADAS. The models are applied to the derivation of sensor parameter constraints for feature-based localisation for urban automated driving and the analysis of performance limitations in AEB systems.
First, environment perception tasks are studied. This comprises the detection of obstacles with a stereo vision sensor and feature-based localisation. Instead of a theoretical top-down analysis, an environment sensor’s inaccuracy can also be empirically evalu ated by comparison to a reference sensor. Advanced off-line approaches can be leveraged to obtain accurate reference data from raw sensor measurements. The potential of such methods, in contrast to an on-line processing under real-time constraints, is studied on the example of a laser scanner sensor.
Second, the models of errors in individual measurements are propagated to the uncertainty in estimates of unobservable dynamic states, for instance motion state variables.
Third, the prediction and risk assessment of traffic situations is considered. Novel expressions for performance bounds on the recognition of semantic driver intentions for long-term motion predictions are derived. Moreover, parametric models of uncertainty in kinematic motion models for short-term predictions are estimated from empirical driving data. Based on these predictions, the risk of an imminent collision can be quantified in terms of criticality measures. A comprehensive approach to probabilistic modelling of these risk metrics is developed. Performance bounds on the timely activation of an AEB intervention are then derived by uncertainty propagation from perception and state estimation to the criticality assessment.