Bayesian filtering for automotive applications

Doctoral thesis, 2015

This thesis is concerned with how data from common automotive sensors can be processed and interpreted in order to support advanced driver assistance systems (ADAS). More specifically, the thesis addresses aspects of object tracking using radar detections, mapping and self-localization for automated vehicles and driver monitoring. In automotive radar tracking, an observed vehicle typically generates multiple detections. This thesis presents a detailed sensor model that adapts to the detection properties of an object by jointly estimating the position of reflection centres and the position of the object. Moreover, the model considers the limited resolution of the radar and evaluation show results close to those achieved with a deterministic vehicle model where the reflecting properties are known. A second contribution to the area of object tracking is a generalization of the well-known cardinalized probability hypothesis density (CPHD) filter to incorporate objects that appear through spawning from existing targets. It is further shown that the generalized filter is tractable for some common birth and spawning models. For automated vehicles, some of the studied problems resemble those traditionally studied in robotics, such as mapping and localization. This thesis presents and evaluates a self-localization solution based on a set of automotive off-the-shelf sensors together with a map that contains lane markings and a simplistic description of radar landmarks. The evaluation shows that this map, in combination with real radar data, provides valuable information to the localization algorithm. With this motivation, a method for estimating more detailed radar maps is derived. The map is modelled by an inhomogeneous Poisson process describing the expected measurements from the static environment as a function of the sensor position. The estimation principle relies on a variational method where the number of landmarks and their respective parameters are found simultaneously. In addition to sensors that observe the vehicle and its surroundings, there are camera-based systems designed to monitor the driver behavior. In the context of driver distraction, this thesis presents a method for driver gaze zone estimation, i.e., estimation of which area the driver is currently looking at, using data provided by such monitoring systems. To improve robustness, the proposed solutions make use of functions that describe the gaze direction based on the head pose and eye closure. It is also shown how these functions can be learnt from data.