Abstract Driver behaviour profiling, specifically in relation to identifying `good' versus `bad' driving patterns, is one of the most challenging problems in mobility data analytics. In this paper, the core task of driver behaviour profiling is addressed at the minimum level of pre-requisites, i.e., GPS-only trajectory data (no accelerometer or other sensors) of very low sampling rate (less than 0.1 Hz). A dynamic temporal resampling algorithm is employed for transforming GPS data into three distinct location-invariant time series, namely speed, acceleration, and turn rate, after map-matching and noise elimination pre-processing steps. A wide range of statistical, time series and spectral methods are implemented as feature functions or `encoders' of various aspects of short-term mobility tracking. In our experimental study, a large real-world trajectory dataset is processed and transformed into such a feature-vector dataset, which is subsequently used in unsupervised training and adaptive category identification for the various driving behaviour `states'. The proposed approach is designed for online/streaming mode and lightweight yet powerful analytics. The results show that such an approach is feasible, despite its challenging context of constraints, providing a data-driven adaptive way to recognizing `normal' vs. `abnormal' driving patterns on-the-fly.

This work was partially supported by the project "Track & Know" (No:780754), which has received funding from the European Union's Horizon 2020 (H2020) programme. http://trackandknowproject.eu