Human Driving Patterns Modeling using Hidden Markov Models and GMM-based Clustering
Reza Haghighi Osgouei, 최승문 · International Conference on Human-Computer Interaction · 2012
In this work the use of two stochastic methods, hidden Markov models (HMM) and Gaussian mixture models (GMM), is presented to construct a reliable model for human driving patterns. After collecting driving signals including velocity and heading of the vehicle using a simulated driving task, translational acceleration and rotational velocity as two most important features are extracted. Then a segmentation method based on detection the local extrema of the velocity and heading signals, divides driving data into a number of segments. By studying histogram plot of the extracted features, it is found a mixture of three and four univariate Gaussian mixture models can be fitted onto translational acceleration and rotational velocity respectively. Considering each mixture component as a cluster, in total we have twelve different clusters in combination. Segmented data are categorized into these clusters; forming training data set for HMMs. The parameters of one HMM for each cluster are optimized using the obtained training data set. The experimental results reveal that the trained HMMs can recognize correct clusters with about 71% accuracy.