Learning driving patterns to support navigation
Thesis DisciplineComputer Science
Degree GrantorUniversity of Canterbury
Degree NameDoctor of Philosophy
Experience is a significant source of knowledge for any human activity. Knowledge about past failures may help to avoid similar failures in the future, while repeating or even improving successes. Driving is a complex and dynamic activity, and the extensive previous experience provides great help when some important decision has to be made quickly. This thesis proposes and demonstrates methods for learning driving patterns and their use in supporting driving navigation tasks. Driving patterns are sequences of events in the traffic system repeating over time. We developed a framework of the driving warning system based on the learned driving patterns. The learning part of the proposed system builds and maintains the model of the traffic system. The system also predicts the most likely future events. The predicted event is compared with the actual event and if/when driver's behaviour becomes significantly different from "usual", appropriate warnings may be generated. We developed a hardware system for data acquisition from navigation sensors. We used off-the-shelf components and a standard software development environment to create an inexpensive but reliable platform for our experiments. Neural networks are computational methods developed to mimic human information processing. In this thesis we present experiments in vehicle movement prediction, using data from a small number of sensors. We tested prediction capabilities of multi-line perceptrons, recurrent and time-delay neural networks. We found that all neural network architectures have good performances for short-term vehicle movement prediction. However, the prediction error becomes significantly higher when predicting events further in the future. Here we also present the new method for driving event recognition based on hidden Markov models. The data from a very limited set of sensors is collected and transformed to observation sequences representing driving events. For each event type we wish to recognize, one hidden Markov model is trained with observation sequences of this type. Sequences representing test events are applied to all models, and the model with the highest probability indicates the type of event for each sequence. Experiments performed show that the proposed method has a very high recognition rate. We developed a system to predict future driving events based on recognized event types and other data collected from sensors. This system was able to successfully predict future events for previously used routes and to detect when an unexpected event has been experienced.