文摘
Travel time estimation is a fundamental measure used in routing and navigation applications,in particular in emerging intelligent transportation systems ITS). For example,many users may prefer the fastest route to their destination and would rely on real-time predicted travel times. It also helps real-time traffic management and traffic light control. Accurate estimation of travel time requires collecting a lot of real-time data from road networks. This data can be collected using a wide variety of sources like inductive loop detectors,video cameras,radio frequency identification RFID) transponders etc. But these systems include deployment of infrastructure which has some limitations and drawbacks. The main drawbacks in these modes are the high cost and the high probability of error caused by prevalence of equipment malfunctions and in the case of sensor based methods,the problem of spatial coverage. As an alternative to traditional way of collecting data using expensive equipment,development of cellular & mobile technology allows for leveraging embedded GPS sensors in smartphones carried by millions of road users. Crowd-sourcing GPS data will allow building traffic monitoring systems that utilize this opportunity for the purpose of accurate and real-time prediction of traffic measures. However,the effectiveness of these systems have not yet been proven or shown in real applications. In this thesis,we study some of the current available data sets and identify the requirements for accurate prediction. In our work,we propose the design for a crowd-sourcing traffic application,including an android-based mobile client and a server architecture. We also develop map-matching method. More importantly,we present prediction methods using machine learning techniques such as support vector regression. Machine learning provides an alternative to traditional statistical method such as using averaged historic data for estimation of travel time. Machine Learning techniques played a key role in estimation in the last two decades. They are proved by providing better accuracy in estimation and in classification. However,employing a machine learning technique in any application requires creative modeling of the system and its sensory data. In this thesis,we model the road network as a graph and train different models for different links on the road. Modeling a road network as graph with nodes and links enables the learner to capture patterns occurring on each segment of road,thereby providing better accuracy. To evaluate the prediction models,we use three sets of data out of which two sets are collected using mobile probing and one set is generated using VISSIM traffic simulator. The results show that crowdsourcing is only more accurate than traditional statistical methods if the input values for input data are very close to the actual values. In particular,when speed of vehicles on a link are concerned,we need to provide the machine learning model with data that is only few minutes old; using average speed of vehicles,for example from the past half hour,as is usually seen in many web based traffic information sources may not allow for better performance.