Research On Short-Time Traffic Condition Prediction Based On Hidden Markov Model

In recent years, with the accelerating pace of urbanization and improving living standard, traveling by car has become the first choice of the public. According to Transportation Ministry, the car ownership in China will reach 20 million by 2020, which will put a higher demand and a huger challenge on the level of the traffic service and the infrastructure construction. With the research on traffic flow, some researchers proposed the theory of Intelligent Transportation System(ITS) that aims to optimize the traffic flow and ease traffic congestion on the basis of existing transportation facilities. And And the ultimate goal is to save energy and reduce pollution and build an harmonious environment. Therefore, much deeper research on ITS has a very important significance for building a much more efficient transportation network.Recentl, short-time traffic condition prediction has been becoming the hot focus of ITS field to timely inform travelers of the traffic condition to plan the best way to travel and select optimal route.Future short-time traffic condition prediction is taken as the research subject in this paper. And the three research directions, namely, the analysis of traffic flow characteristics, the building and analysis of prediction model and the evaluation of prediction results, are given. The time-varying and nonlinear characteristics on timeline are further understood through the analysis for the traffic data sequences. It can be seen that the traffic flow is an time-varying stochastic process.Even though the availablity of some prediction models or methods, they are quantitative and deterministic with only considering the static traffic information, while ignoring the dynamic information. For the reason, Hidden Markov Model are used for the traffic states prediction study.Firstly, the Traffic Hidden Markov Model(THMM) is studied based on the statistical analysis of the data sets. The hidden states set is obtained through the combination of the two kinds of discretion grades through discretizing the average value and the contrast of the data sequences in the prediction window. And the observed states correspond to the discretization value intervals of the first measured value in prediction window. Consequently,the hidden and observed states set are given.Secondly, the model parameters including states transtion matrix, emission matrix and the initial probability distribution are given by training the model using the EM algorithm based on the training data sets.Finally, on the base of the well-trained model parameters, the Hidden Markov Model is utilized to predict the future short-time traffic states combined with different kind traffic flow parameter and its corressponding training data set under varying prediction sequences length. And the prediction results are analyzed.