Trajectory Annotation by Discovering Driving Patterns

Abstract

The ubiquity and variety of available sensors has enabled the collection of voluminous datasets of car trajectories that enable analysts to make sense of driving patterns and behaviors. One approach to obtain driving behaviors is to break a trajectory into its underlying patterns and then analyze these patterns (aka segmentation). To validate and improve automated trajectory segmentation algorithms, there is a crucial need for a ground-truth against which to compare the results of the algorithms. To the best of our knowledge, no such publicly available ground-truth of car trajectory annotations exists. In this paper, we introduce a trajectory annotation framework and use it to annotate a real-world dataset of personal car trajectories. Our annotation methodology consists of a crowd-sourcing step followed by a precise process of expert aggregation. Our annotation identifies segment borders, and then labels the segment with its type (e.g. speed-up, turn, merge, etc.). The output of our project is a dataset of annotated car trajectories (DACT), and is publicly available for use by the spatiotemporal research community at https://goo.gl/XgsxyJ.