A Data-driven Map-matching Method For Driving Trajectory Data

Satellite positioning data are the data generated by the satellite positioning system to determine the current geographical location of a target,and are widely used in the fields of military,people's livelihood and scientific research.However,there is inevitably a certain extent of deviation between the satellite positioning data and the actual position of the target,which brings some troubles to the practical application of the satellite positioning data.Map matching is a process to improve the accuracy of satellite positioning data by using map data.It is difficult for model-based map-matching methods currently used to deal with the condition of change in data types.However,the condition of change in data types is fairly common: research shows that the introduction of new types of data can effectively improve the accuracy of map matching,and on the other hand there is inevitably data loss in practice.In order to solve this problem,this dissertation proposes data-driven map-matching methods which train a machine learning model to cope with the change in data types.These methods have great flexibility and high accuracy.Our works include the following three aspects:(1)Visual analysis method of map-matching data: data quality directly determines the performance of data-driven methods.The data used for map matching is numerical data which are so abstract that human beings are not able to understand and analyze intuitively,therefore,it is difficult for humans to determine whether the data used for map-matching is reliable.To solve this problem,we proposed a visual analysis method.This method designs different visualization methods for road network data,GPS data,training data and other data used in map matching respectively so as to present these data to users in the form of graphics,thus users can understand and analyze these data to find problems of the data and then improve these problems.As a result,the reliability of the data is ensured.Our case studies show that our visual analysis method can find problems in the training data and the causes of these problems,and help users improve the reliability of the used data.(2)Local map-matching method based on ranking learning: the problem of local map-matching is neither a classification problem nor a regression problem of supervised machine learning,and therefore cannot be solved by the commonly used machine learning model.In order to solve the problem,this dissertation regards mapmatching problem as ranking problem,and proposes a local map-matching method.This method first uses a Gaussian distribution to select initial matching results,named candidate roads,then uses the ranking learning method to train the neural network designed by us to learn a discriminator of candidates road for selecting the most possible road,and finally uses the trained discriminator to select the matching results from the candidate roads.Experimental results show that our local map-matching method can effectively deal with the change of data types,and has better accuracy and execution speed.(3)Global map-matching method based on DNN-HMM hybrid model: due to without global features in matching,the accuracy of local map-matching methods is lower than that of global map-matching methods.In order to improve the accuracy of our local map-matching method,we propose a DNN-HMM model to realize a datadriven global map-matching method.This hybrid model proposes a conversion formula in order to convert the unexplained local map-matching results generated by our local method into a normalized state-observation probability of HMM,a method to get corresponding state-transition probability of HMM,and a method to figure out global matching results.Experiments show that our global method can not only deal with the change in data types effectively,but also achieve the state of the art of map matching methods in the accuracy.