Feature extraction and classification of sensor signals in cars based on a modified codebook approach

In this paper we indicatively address the problem of classifying sensor signals transmitted on the Controller Area Network (CAN bus) inside cars. The aim of this work is to solve the problem of dealing with large amounts of CAN bus data, finding out the semantics of the most important signals inside CAN buses in cars, independently of the cars' model and manufacture. For the purpose, we are aiming at classifying vehicle CAN bus data into basic sensor signals like speed, brake, steering-angle and throttle that assume a prior role in finding driving behaviors. In particular, our approach is starting with two feature extraction methods, a handcrafted feature selection (HFS) and an automated feature selection (AFS). Our automated method is based on a codebook approach, which extracts histogram-based features characterizing the shape of signals. Since some of the signal data are behaving very similar but with different semantics, in order to gain a better result, not only a one-vs-one classifier but also a multi-class classification algorithm has been applied. The result of our approach is a comparison between the applied methods, which validates the use of innovative AFS.