Clustering of Driving Scenarios Using Connected Vehicle Datasets
Driving encounter classification and analysis can benefit autonomous vehicles to efficiently achieve a more smart decision. This paper presents an unsupervised learning framework to classify a wide range of driving encounters which compose of a pair of vehicles' GPS trajectories ordered by time. First, we develop five specific approaches, through integrating deep autoencoders with distance-based measures, to learn underlying representations of driving encounters in terms of both shape and distance, and we thoroughly compare and evaluate the performance of the five approaches. We then apply k-means clustering to categorize driving encounters into distinguishable groups based on the learned representations. Our proposed unsupervised learning framework for driving encounter classification is finally validated based on 2568 naturalistic driving encounters from connected vehicle datasets.
READ FULL TEXT