Bayesian Inference for Automotive Applications
Environment perception is an important aspect of modern automated systems. The perception consists of fusing information from different sensors to estimate variables which provide a description of the scene. The main contributions of this thesis are in multiple object tracking and mapping, and nonlinear filtering methods for road geometry estimation, which are important topics in the design of advanced driver assistance systems and self-driving cars.
Road geometry estimation is required in many advanced driver assistance systems. In this thesis, we utilize non-linear filtering methods to perform long-range road geometry estimation by fusing measurements from
different sensors. In this context, the process model describes the time evolution of the road state and the measurement model establishes the relationship
between the sensor measurements and the road state. We parameterize our road model such that the time evolution of the state follows the manner by which roads are built and design a filtering algorithm to estimate it.
Sensor maps provide a description of the environment as seen through the lens of a sensor. These maps describe the measurement distribution of an environment as a function of the sensor position and are used to perform localization. In this thesis, we model the prior of a radar map by a Poisson process which allows us to incorporate the uncertainties in the number of landmarks, their states and data association hypotheses, into the model. We derive the exact theoretical batch multi-object posterior density of the map and use Gibbs sampling method to approximate the posterior.
Extended object tracking is an important problem in the context of multiple-object tracking which arises, for example, in tracking using automotive radars. We present a Poisson multi-Bernoulli mixture conjugate
prior for multiple extended object tracking. Since the resulting posterior is intractable, approximations are required to obtain a feasible algorithm. We present two tractable solutions, one based on the full PMBM posterior and one based on approximating the PMBM by a PMB process. Multiple practical challenges arise in connection to the PMB filter, to which we provide pragmatic solutions that yield an efficient and tractable algorithm.