| With the expansion of urban scale and the geographical division of urban functions in China,residents’ travel modes have become more complex and varied both in time and space,causing great challenges to urban traffic management and passenger flow forecasting tasks.The essence of traffic demand is to meet people’s need for daily travel.Deeply exploiting the mechanism of urban traffic system and revealing the spatial-temporal characteristics of residents’ travel patterns can provide important support for the construction of high quality and efficient urban transportation system,which makes sense.Traditionally,residents’ travel information is collected by household travel survey methods such as paper questionnaire and telephone interview.For a long time,there exists many data quality and collection efficiency problems,such as data distortion due to participants’ subjective memory,complicate and complex investigation organization,long data update cycle and so on,making it difficult to support sophisticate travel analysis for modern traffic plans.With the advent of the big data era and the Internet age,it is promising to track and observe residents’ travel patterns using individual smart phone data.Smart phone sensor-based survey method which uses sensing units such as GPS,accelerometer,gyroscope,WiFi et al.to collect temporal-spatial travel information and travel status data,can better depict individual’s travel behavior.The collected data are also more objective and accurate.It is becoming a hot topic in the field of traffic information acquisition.Based on the research of data collection,data analysis and data mining,this paper explores new methods for extracting individual trip chain information using mobile sensor data.The main research contents and conclusions are as follows:1)The smart phones sensor modules and data acquisition functions are analyzed,a smart phone sensor data collection APP is developed.The APP connects smart phone GPS,accelerometer,gyro,WiFi,etc.module ports,can collect individual spatio-temporal travel information(travel time,longitude and latitude,direction angle,etc.)and travel status information(speed,3D acceleration,3D angular velocity,positioning accuracy,station ID etc.)in real time.A database is also built to manage and storage data efficiently and safely.2)The data quality influencing factors of different technology and environment are studied and the field data collection work is completed.Data characteristics change a lot under different experiment environment.Considering urban population and traffic environment in China,this paper collects experiment data under various designed test constraints,the data sampling frequencies include 1s、2s、3s、4s、5s、10s,20s、30s、60s、120s;Travel modes include 10 combination patterns constituted by walk,bike,bus,car,subway,etc;Travel purposes include 8 commutes and non-commutes travels,such as go to work,go to school,visit friends,see doctor,go shopping etc;Travel time includes holidays,weekends,working days,morning rush hour,evening rush hour and off-peak hours;Traffic conditions include free flow,generally congested and heavily congested patterns.3)Smart phone sensor data content and format are analyzed,using map matching technology and 3d-visualization technology,the data fluctuation characteristics and change rules are further explored,the expression level and expression method of multi-source smart phone sensor data for individual travel behavior are also researched.Depend on the analysis of data characteristics and data change rules,attributes for model construction are explored under different travel environment.Therefore,the proposed algorithms are more portable and compatible.Results show that the time-space density of positioning points and the instantaneous travel speed index are applicable to trip end identification and mode transfer point identification respectively,the average speed,the maximum speed,the variance of velocity,and the variance of acceleration per minuter are the best input attributes for travel mode identification.4)The overall technical framework,the phased processing process and the key points in each phase for individual trip chain information extraction are presented.The process mainly contains three steps: First,trip ends are indentifed by proposing the spatial-temporal clustering algorithm,travel status detection algorithm and error optimization algorithms,one day or multi-day trip data are devided into subsegments and each subsegment contains only one trip purpose(Identification of trip OD);Second,the wavelet transform modulus maximum algorithm is introduced for mode transfer point identification,each subsegment are furthered devided into segments which contains only one travel mode(Identification of mode transfer point);Last,machine learning methods(such as neural network,support vector machine,bayesian network,random forest)are explored to detect trip mode,a GIS map matching method is proposed to solve the bus and car mode detection problem.Results show that trip end detection accuracy reaches 86% and the dwell time detection errors are all within 150 s.The rate of mode transfer point detection reaches 95%,and the average error of mode transfer time detection is less than 1 minute;Trip mode detection accuracy reaches 90%.5)The model parameters were calibrated by 10-fold cross validation,based on the thought of experimental engineering,the construction of individual trip chain parameter extraction model and the results feedback optimization method are researched.The empirical evaluation and sensitivity analysis tasks are conducted,trip mode detection accuracy under different sampling frequencies,traffic conditions and model algorithms are evaluated.Results show that the best data sampling frequency is 5s,and it should be better smaller than 20 seconds in practice;Algorithm significantly affects the effect of travel mode detection,machine learning methiod cooperated with GIS map maching method achieves better results than only use maching learning method,bus and car mode detection accuracy achieve nearly 20% improvement;Traffic condition has certain influence on trip mode detection,but the GIS map maching method can reduce the detection error. |
