Vehicular traffic density estimation via statistical methods with automated state learning

Abstract

This paper proposes a novel approach of combining an unsupervised clustering scheme called AutoClass with Hidden Markov Models (HMMs) to determine the traffic density state in a Region Of Interest (ROI) of a road in a traffic video. Firstly, low-level features are extracted from the ROI of each frame. Secondly, an unsupervised clustering algorithm called AutoClass is then applied to the low-level features to obtain a set of clusters for each pre-defined traffic density state. Finally, four HMM models are constructed for each traffic state respectively with each cluster corresponding to a state in the HMM and the structure of HMM is determined based on the cluster information. This approach improves over previous approaches that used Gaussian Mixture HMMs (GMHMM) by circumventing the need to make an arbitrary choice on the structure of the HMM as well as determining the number of mixtures used for each density traffic state. The results show that this approach can classify the traffic density in a ROI of a traffic video accurately with the property of being able to handle the varying illumination elegantly.