Identifying Truck Stops from a Large Stream of GPS Data via a Hidden Markov Chain Model

This paper proposes and shows the application of the Hidden Markov Model (HMM) to identify truck trip segments and extract activity and non-activity stops from large scale truck GPS data while accounting for the spatiotemporal properties of GPS points, and hence overcoming limitations of existing clustering practices in freight studies. Individual trajectories and different classes of stops were extracted from four years of raw GPS data of Australian high-performance trucks, totalling more than 71 million records. The framework presented in this paper is easy to re-apply and can be transferred to other large datasets, without requiring secondary data sources such as drivers' logbooks to infer the trip ends. The results unveil three distinct type of stops across trajectories (stops due to traffic congestion, activity, and non-activity stops), and their duration, providing empirical insights into the diverse aspects of long-haul freight transportation. The validation showed satisfactory results where the modelled trajectory episodes yielded similar results to the known trip trajectories, showing a 97% accuracy rate in identifying the activity stops. The validation showed that the percentage of activity stops that the model incorrectly assigned to any other state was 3% (false negative error), and the percentage of instances that mistakenly identified as activity stops by the model was only 1%.