Prediction of driving behaviors in intersections based on a supervised dimension reduction considering locality

Prediction of driving behavior has been regarded as one of the important issue to realize the next generation of advanced driver assistance systems. However, prediction of driving behaviors is also difficult issue, because the distribution of each driving behavior seems to be not unimodal but multimodal due to its intrinsic complexity and lack of a well-established segmentation method. When we consider to predict driving behaviors with a supervised dimension reduction method and hidden Markov models (HMMs), the multimodal structure of observed distributions should be preserved since they can contain information regarding these behaviors. We therefore propose to combine HMMs with local Fisher discriminant analysis (LFDA) that can maximize the between-class separatability and preserve within-class multimodality. We evaluated the performance of the HMMs with LFDA in predicting actual driving behaviors, and compared its performance with those of with conventional Fisher discriminant analysis and with no dimension reduction. As a result, the LFDA based method showed the best prediction accuracy among the all methods.