Vehicle Classification from Low Frequency GPS Data

摘要

Inferring the type of vehicles on a road is a fundamental task within several applications. Some recent works have exploited Global Positioning System (GPS) devices and used classification of GPS traces to tackle the problem. Existing approaches based on GPS data make use of GPS trajectories sampled at high frequency (about 1 sample per second), but GPS trackers currently installed on public and commercial fleets acquire GPS positions at lower frequency (about 1 sample per minute). In this paper, we target the more challenging scenario of low frequency GPS data, which has not been tackled yet in the literature, and explore how this kind of data can be used to effectively categorise vehicles into light-duty and heavy-duty. We define several distance-, speed-, and acceleration-based features, inspired by the literature on related problems like travel mode detection, and add novel features based on road type. Features are aggregated over a GPS track with several aggregation functions. We identify the most effective combinations of features and aggregation functions with a data-driven approach, by applying Recursive Feature Elimination in a cross validation framework. Furthermore, we combine predictions of all tracks of a vehicle to boost classification performance. Experimental results on a large dataset show that the selected features are indeed effective and that the high and low frequency GPS scenarios greatly differ in terms of relevant features.