Filtering and modelling for automotive safety systems

This thesis makes five important contributions to the development of an automotive safety system: filtering algorithms, three modelling frameworks concerning the usage of radar detections in tracking, vehicle motion, and decision-making for intervention decisions, and finally the implementation architecture. In the filtering context, we have developed a new sigma-point method for estimating the moments of a transformed Gaussian random variable. These estimates are derived from analytical expressions and are based on evaluations of the transforming function. The method is applied to the moment estimation task in a Gaussian filter and the resulting algorithm is denoted the marginalised Kalman filter (MKF). Compared to traditional radar models, ours is specifically designed for vehicle radars, which often yield several measurements from each object. These measurements can provide useful information, such as vehicle orientation, if they are accurately modelled. We introduce a tracking filter using such a sensor model, and show how the complex data association problem can be facilitated by merging similar hypotheses into groups. The presented vehicle motion model includes the control input from the driver. Uncertainties regarding, e.g., driver style, are formally treated with increased prediction accuracy as a result. Similar to this model, the third framework also takes the driver into consideration by allowing interventions only when the driver is believed to accept them. Our evaluations indicate an increased benefit in collision avoidance systems --- particularly in traffic situations where the future trajectory of another road user is hard for the driver to predict. Finally, we present a modular functional design for implementing a real-time data fusion system. We conclude that a tracking system, using modern estimation techniques, is well suited for sensor data fusion in an automotive environment