Nonlinear Filtering Methods in Road Geometry Estimation

Licentiate thesis, 2014

Advanced driver assistance systems require information about the traffic scene of which the road geometry is an important part. In this thesis we use nonlinear filtering methods to estimate the shape of the road ahead of a host vehicle by fusing measurements of lane markings obtained from a camera and measurements of moving vehicles and barriers obtained from a radar-camera fusion system. We focus on highway scenarios where road geometry information up to long distances (possibly longer than 100m) is required, due to high speeds of the vehicles. We develop a clothoid-based road model which describes the shape of a highway-type road up to 200m ahead of the host vehicle. We describe the road geometry estimation problem by two models. A process model that describes the time evolution of the road state and a measurement model that establishes the relationship between the sensor measurements and the road state. We parameterize our road model such that the parameters are closely tied to the physical interpretation of roads, hence, the time evolution of the state follows the manner by which roads are built. We develop measurement models and strategies to make use of the available information in a filter. The first model describes how the measurements of lane markings are related to the road state. The second describes how headings of the leading vehicles provide information about the road. Leading vehicles can either follow their lane or make a lane change, and we are mainly interested in using the heading data from vehicles that are following their lane. We present two approaches to handle the possibility that the leading vehicles may not be following their lane. One where we develop a probabilistic framework based on multiple-model filtering, and another where we use an outlier detection algorithm. The third model describes how the barrier measurements relate to the state vector. Based on these measurements we use the Hough transform to form hypotheses indicating which measurements are generated by the barriers. We present probabilistic models for the measurements of each barrier which enables us to make use of this information in conventional filtering algorithms.