The increasing availability of large-scale Global Positioning System (GPS) data stemming from in-vehicle embedded terminal devices enables the design of methods deriving road network cartographic information from drivers' recorded traces. Some machine learning approaches have been proposed in the past to train automatic road network map inference, and recently this approach has been successfully extended to infer road attributes as well, such as speed limitation or number of lanes. In this paper, we address the problem of detecting traffic signals from a set of vehicle speed profiles, under a classification perspective. Each data instance is a speed versus distance plot depicting over a hundred profiles on a 100-meter-long road span. We proposed three different ways of deriving features: the first one relies on the raw speed measurements; the second one uses image recognition techniques; and the third one is based on functional data analysis. We input them into most commonly used classification algorithms and a comparative analysis demonstrated that a functional description of speed profiles with wavelet transforms seems to outperform the other approaches with most of the tested classifiers. It also highlighted that Random Forests yield an accurate detection of traffic signals, regardless of the chosen feature extraction method, while keeping a remarkably low confusion rate with stop signs.