A Hidden Markov Model for Vehicle Detection and Counting

To reduce roadway congestion and improve traffic safety, accurate traffic metrics, such as number of vehicles travelling through lane-ways, are required. Unfortunately most existing infrastructure, such as loop-detectors and many video detectors, do not feasibly provide accurate vehicle counts. Consequently, a novel method is proposed which models vehicle motion using hidden Markov models (HMM). The proposed method represents a specified small region of the roadway as 'empty', 'vehicle entering', 'vehicle inside', and 'vehicle exiting', and then applies a modified Viterbi algorithm to the HMM sequential estimation framework to initialize and track vehicles. Vehicle observations are obtained using an Adaboost trained Haar-like feature detector. When tested on 88 hours of video, from three distinct locations, the proposed method proved to be robust to changes in lighting conditions, moving shadows, and camera motion, and consistently out-performed Multiple Target Tracking (MTT) and Virtual Detection Line(VDL) implementations. The median vehicle count error of the proposed method is lower than MTT and VDL by 28%, and 70% respectively. As future work, this algorithm will be implemented to provide the traffic industry with improved automated vehicle counting, with the intent to eventually provide real-time counts.

I. Introduction

Intelligent transportation systems (ITS) are increasingly reliant on sophisticated applications of sensing and control, including live real-time video analytics. Reliable real-time vehicle traffic information is essential for an effective implementation of ITS to ease congestion and reduce delays. A desirable measurement for predicting demand and efficiently controlling traffic light timing and routing is to count the total volume of vehicles over time through each intersection leg [1], [2]. For full integration into an effective intelligent transportation system, reliable volume counts should be provided online in real-time.