Reducing the negative effect of defective data on driving behavior segmentation via a deep sparse autoencoder
Hailong Liu, Tadahiro Taniguchi, Kazuhito Takenaka, Yuusuke Tanaka, Takashi Bando · 2016
Analyzing driving behavior data is essential for developing driver assistance systems. Statistical segmentation is one of the important methods to realize the analysis. Driving behavior data actually include undesirable defects caused by external environment and sensor failures. Defects in the data cause a huge negative effect on the segmentation. In this paper, we showed that a feature extraction method based on a deep sparse autoencoder with fixed point (DSAE-FP) could reduce the negative effect of defective data in a driving behavior segmentation task. In the experiments, we used sticky hierarchical Dirichlet process hidden Markov model to segment the driving behavior. We compared the segmentation results using hidden features extracted by DSAE-FP and other comparative methods. Experimental results showed that segmentation results of non-defective dataset and defective dataset turned out most similar when DSAE-FP was used.