Study On The Individual Travel Pattern Recognition Based On Smartphones

From the traditional questionnaire passively enquiring useful information to the GPS device or smartphone proactively collecting data,the technology innovation of travel behavior survey has been facing a new opportunity.However,how to deeply mine the travel features and precisely recognize the travel patterns including dynamic travel and static activity based on GPS trajectory data has become the critical premise of smartphone application to the travel-activity behavior research.The thesis obtains individual GPS trajectory data by conducting survey with smartphone application,then detects anchors from the signal missing segments and the complete segments,and divides the scattered track points into travel chain.Subsequently,some related features of anchors and travels are respectively derived to recognize the travel mode and trip purpose with the help of geographic information system and application of machine learning algorithm.Firstly,the principle and method of GPS data collection with smartphone and data structure have been introduced,followed with the pretreatment steps,for instant,filtering the noise from the perspective of data accuracy,normalspeedrange,latitude and longitude scope,thenconvertingcoordinates,to lay the foundation for the pattern recognition.Secondly,according to the interval,the daily trajectory has been divided into the signal missing segments and the complete segments.The former segments are checked whether an anchor exerts by calculating the interval and distance of adjacent endpoints,while the latter are recognized with a new temporal-spatial clustering algorithm learned from DBSCAN.Afterwards,the results have been compared with the questionnaire feedback data to optimize the parameter threshold and ensure the anchors and travels with stop times,duration or other evaluation indictors.Thirdly,the travels are cut apart into trip legs with the help of identified walk segments.After the statistical features like velocity,acceleration,heading change rate,and the travel features,such as duration,distance,stop times and so on are derived from trip legs,subsequently classification and regression tree,neural network,random forest are applied to recognize travel mode.It suggests that random forest performed better and gets the comprehensive accuracy of 89%.Furthermore,a contrast situation is established by not considering e-bike,which discovers that e-bike has a great effect on bicycle and bus detection,in addition,the interference between e-bike and bicycle,and the interference between bus and car become the critical factors on recognition performance.While the other contrast situation ignores the travel features,result suggests that taking the travel features into consideration contributes to improving accuracy of four confused modes on some extent.Finally,according to the frequency distribution of start time and duration of anchors,temporal features are derived.Meanwhile,it is necessary to search and count the POI numbers of each activity category in the area surrounding anchor center in a radius of 50 meters and 300 meters by the use of web map API,and calculate the entropy to be regarded as the spatial features.Classification and regression tree,neural network,and random forest algorithm are applied to recognize the trip purpose.It shows that the accuracy ultimately reaches 80 percent by random forest algorithm which has a better performance than other algorithms.The result indicates that the commuting has highest accuracy,while the shopping and recreation are lowest.The interference between eating out and shopping becomes the main factors affecting the recognition.