Action pattern recognition based on Action phases clustering

Abstract

Current approaches to identifying driving heterogeneity often struggle with accurately deciphering fundamental patterns inherent in driving behaviour. The concept of Action phase is proposed to capture underlying driving characteristics with physical meanings. This study further recognises Action patterns by clustering extracted Action phases. A Resampling and Downsampling Method (RDM) is first applied to standardise Action phases length. Using features selected by Principal Component Analysis (PCA), two clustering algorithms, i.e., Agglomerative clustering with dynamic tree cut and X-means clustering, are utilised to group Action phases with high similarity. Six Action patterns named “Catch up”, “Fall behind”, “Follow behind”, “Speed up”, “Slow down”, and “Hold speed” are finally recognised based on clustering results. Further statistical analyses demonstrate that velocity and time headway exhibit higher importance than other variables in characterising driving behaviour. The methodology and findings presented in this study offer a nuanced approach to interpreting driving behaviours, which can help to enhance the accuracy of driving heterogeneity identification.