Abnormal driving behavior detection using a linear non-Gaussian acyclic model for causal discovery

Abstract

This paper proposed a causal discovery method for driving behavior signals representing explicit driving actions to gain a better understanding in driving actions and to apply them to improve advanced driver assistance systems. The first- and second-order statistics have only correlation information and hence they cannot tell anything on causality. In other words, causality appears in higher-order statistics. To utilize such information, we apply a linear non-Gaussian acyclic model (LiNGAM) that assumes non-Gaussian observation noise to the driving actions. Non-Gaussian assumption makes it possible to use ICA to estimate the mixing matrix of observed variables. The mixing matrix can transform to a triangular matrix that expresses causality. This method is called ICA-LiNGAM. ICA-LiNGAM can extract causality in time-series of vectors. Driving actions are time-series of vectors that play an essential role in the safe driving. To obtain driving data, we carried out experiments in which subjects drove a freeway without any instruction except for its route. In our method, both instantaneous and lagged causal influences have been considered by using the Structural Vector Autoregression (SVAR) to the continuous time series data. SVAR model is a generalization of LiNGAM model so the assumption of independence and non-Gaussianity is ensured for ICA-LiNGAM analysis. The conventional LiNGAM can be regarded as a special case of SVAR when the autoregressive order is zero. In the experiment ICA-LiNGAM worked well with SVAR model and the analysis result showed a clear causal relationship between lagged variables of driving action signals. The experiment shows that ICA-LiNGAM can find out both the obvious and latent causal relations between driving behaviors. © Springer-Verlag 2013.