Driving Maneuver Classification: A Comparison of Feature Extraction Methods

Driving maneuver classification has received increasing attention in recent years. Early work focused on car-based sensor systems, but recently the use of smartphone-based sensors has been increasingly favored. For a driving maneuver classification system, feature extraction often plays an important role. Previous studies have proposed various feature extraction methods for classifying driving maneuvers, however, a direct comparison of feature extraction methods using various data sets is missing. In this paper, we systematically compare three window-based feature extraction methods for driving maneuver classification: statistical values and automatically extracted features using principal component analysis and stacked sparse auto-encoders. Specifically, all sensor information from each data set is first segmented into windowed signals after preprocessing. Then, the three feature extraction methods are applied to those windowed signals. Finally, extracted features are fed into a random forest classifier. Maneuver classification performance is evaluated on three different data sets, demonstrating weighted classification F1-scores of 68.56%, 80.87%, and 87.26%. For all three data sets, statistical features achieve the best performance.