Identification of Driving Scenarios and Driving Styles Using Machine Learning
Andreas Nord · 2020
The data-based verification of autonomous driving functionalities requires the detection of driving scenarios and driving styles. Driving scenarios define the tasks that the functionalities have to master, whereas the driving style has an impact on the traffic and thus the recorded traffic scenarios. In this thesis we consider two machine learning models which extract driving scenarios and driving styles respectively, and compare their performance on differentiating drivers from a data set with over 350 000 km of driving. The first approach uses Hierarchical Dirichlet Process Hidden semi-Markov models to segment the data into highly interpretable primitive driving scenarios, from which features are extracted to classify drivers. The second approach is a deep learning classifier which utilizes a combination of a sequence model and an auto-encoder. The deep learning classifier outperforms the segmentation based classifier and achieves high accuracy on classifying multiple drivers. This indicates the existence of sufficiently different driving styles to suggest a benefit in assigning multiple drivers for data collection in order to increase the variation in the experienced scenarios. While the segmentation approach is underperforming on the classification task, it can be utilized to replace a rule-based extraction of simple driving scenarios. We discuss the extraction of complementary data characteristics from the two approaches and their relative advantages.