Resident Trip Feature Research Based On Taxi Trajectory Data Mining

Residents' travel behavior analysis plays an essential role in integrated transportation planning and urban construction planning,also providing an important basis for city traffic policy formulation.Travel data used to be collected by household interview and questionnaire survey with high mistaken reporting rate and time-wasting,which fails to meet the needs of modern society.With rapid advances in Geography Information System(GIS)and wide deployment of Global Positioning System(GPS),a large number of individual moving trajectory data is widely stored,providing a new thought for the analysis of residents' trip.Based on the real-world GPS traces of 12000 taxi in Xian over a month,the paper focus on discovering residents trip features from taxi GPS trajectory,which were applied to travel behavior analysis,travel hotspots extraction and area land-use classification.The paper proceeds as follows.To begin with,MapReduce computing framework based on cloud computing was used for large-scale GPS data ranking and trajectory extraction.The realization of data cleaning and map matching work were introduced.Secondly,OD information was extracted from GPS data and various features of different dimension were designed to analysis residents' different travel rules between workday and holiday,such as average travel numbers,average travel time and average travel distance.Visualization measures were used to express spatial distribution of residents' travel.Besides,an improved DBSCAN algorithm is proposed to solve the problem of traditional DBSCAN algorithm,which is sensitive to parameter and fails to limit the cluster size.The improved algorithm chooses parameters adaptively based on average neighbor density and the area constraint of hotspots is given.Clusters beyond the specified size will be split to small clusters,then,hotspots with suitable size were clustered from a great deal of OD points.Finally,flow timing characteristic of different dimensions were extracted to describe the relationship between social function of regions and flow patterns.A semi-supervised learning classification algorithm combined with uncertainty sampling active learning method is proposed and applied to hotspots land-use identification.The hotspots were successfully divided into six categories: stations,scenic areas,commercial areas,residential areas and schools.Experimental results reveal that taxi trajectory data reflect urban residents trip spatial distribution regulation very well.The improved DBSCAN algorithm can be used to cluster residents travel hot areas with reasonable acreage,avoiding the problem of size restriction clustered by traditional algorithim.The social functions of the regions were identified with flow characteristics and the finer features got a better classification results.The semi supervised classification algorithm combined with active learning method could reach high classification accuracy with just a few regions labeled.