Driving Behavior Evaluation in Real-World using Time Series Clustering Method

– Similarity Calculation Considering Shape and Temporal Components in Time-Series Data –

Abstract

There are drivers who are prone to accidents. To find drivers with high crash risk, in-cabin systems are required to monitor and evaluate driver’s daily driving behavior. Recently, a large of driving data has been recroded, but it is difficult to label them. Therefore, in this study, we propose to develop a time series clustering based driving behavior evaluation system, which does not require pre-defined rules or labels. A k-means method with Dynamic Time Warping (DTW) and DTW Barycenter Averaging (DBA) is used to cluster time series data. Conventional DTW was improved to deal with temporal differences between time series. The experimental resuls show that the proposed system could effectively detect abnormal driving behaior, and is more effective in actual driving conditions than rule-based methods.