Driving context classification using pattern recognition

Abstract

The performance of a vehicle system is to a large extent dependent on the driving context, such as the road infrastructure, in which the vehicle is operating. In order to achieve improved performance, di erent vehicle system applications may need to take driving context parameters into account. In this thesis, we develop a pattern recognition framework that classi es driving context based on data recorded by vehicles (speed, steering wheel angle, etc.) in a naturalistic setting. We train the framework on a large data set of vehicle data annotated with map attributes from a map database representing driving context. An inventory is made on the map attributes, nding two kinds of global-scale driving context classes to classify: (1) whether a vehicle is driving in a city or not, and (2) the functional class of the road the vehicle is driving on. We then review four pattern recognition models: Logistic Regression, SVM, Hidden Markov Model, and a simple Baseline model, comparing their ability to classify (1) and (2). We nd that all models reach similar overall prediction accuracies, ranging between 76 % - 80 % for classi cation task (1) and 84 % - 86 % for task (2), but that the models di er slightly with respect to per-class prediction accuracy.