An Image-Based Approach for Classification of Driving Behaviour Using CNNs
In this work we present an approach for the classification of driving behaviour using Convolutional Neural Networks (CNNs), based on measurements that have been obtained by the internal CAN-bus of the vehicle. As is the case with different driving behavio
- PDF / 752,685 Bytes
- 9 Pages / 439.37 x 666.142 pts Page_size
- 9 Downloads / 177 Views
2
Department of Computer Science and Telecommunications, University of Thessaly, Lamia, Greece {espyrou,ivernikos}@uth.gr General Department, University of Thessaly, Lamia, Greece {msavelonas,sk}@uth.gr
Abstract. In this work we present an approach for the classification of driving behaviour using Convolutional Neural Networks (CNNs), based on measurements that have been obtained by the internal CAN-bus of the vehicle. As is the case with different driving behaviours, CAN-bus sensor data reflect the driving patterns associated with different types of vehicles. The experimental evaluation is performed on a real-life dataset composed by measuring 27 attributes, for 4 different car types, namely vacuum, car, truck and garbage truck. These features are processed to form pseudocolored images, capturing both temporal and qualitative features of parts of routes. For classification, we use a deep CNN architecture. Results indicated an accuracy of 91% and increased performance compared to other neural network-based approaches. Keywords: Convolutional neural networks behaviour · CAN-bus measurements
1
· Deep learning · Driving
Introduction
Controller area network (CAN)-bus standards using on-board diagnostics (OBD) ports, as well as telematics sensors, have been shown to provide a signature of driving behaviour by Meiring and Myburgh [1]. Several intelligent approaches analysing such data have been proposed in the last decade. In [2], statistical features of various driving events have been derived and used in the context of either unsupervised learning (k-means clustering) or supervised learning (SVMs), so as to distinguish different drivers. In [3], several statistical features of 3-axis accelerometer signals where calculated and used to classify driving behaviour as normal or aggressive. These features include central and dispersion characteristics, histogram moments, Kendall’s tau rank correlation coefficient, covariance between each pair of acceleration signals etc. Using sequential c The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 E. G. Nathanail et al. (Eds.): CSUM2020, AISC 1278, pp. 263–271, 2021. https://doi.org/10.1007/978-3-030-61075-3_26
264
E. Spyrou et al.
forward feature selection and k-nearest neighbours (k-NN) classification, they maintained 7 features and achieved 100% accuracy. However, the authors note that the driving samples comprising their dataset might be too easily separable. In [4], thresholds associated with events such as acceleration, braking and turning where defined and used for event detection, aiming at classifying driving behaviour on data acquired by mobile phone sensors. The same group introduced UAH dataset [5], acquired by means of mobile phone sensors. This dataset comprises route samples of normal, aggressive and drowsy driving behaviour, under various conditions (motorway or secondary road, six different drivers etc.). The UAH dataset facilitates research in driving behaviour, although mobile phone sensors are less acc
Data Loading...
