Research Of Travel Mode Recognition Methoed Based On Cellphone Accelerometer Data

Acquiring household travel information accurately and effectively can be very helpful for traffic management apartment to carry out traffic facility construction, adjust road resource allocation and raise travel efficiency of residents. And with the acceleration of urbanization, urban traffic status changes at every moment, traditional household travel survey becomes lame because of long investigation period and high cost. Cellphone accelerometer technology rises as smartphone becomes popular, and its features of instantaneity, dynamic nature and sustainability have raised attention of some scholars in transportation industry, it has been a trend to apply cellphone accelerometer data to assist household travel information survey.This paper used cellphone accelerometer data for travel mode recognition, evaluated mode recognition algorithms using training data. Firstly, this paper studied four common mode recognition algorithms, introduced the definitions, principles, merits and demerits of those algorithms briefly.Then designed an experiment to collect cellphone accelerometer data which contained two travel routes and several travel modes, an app called GPSurvey was used to collect data, meanwhile travel log was kept for comparison. When obtained, the original data was analyzed preliminarily and its features were extracted to lay the foundation for the following mode recognition.And chose data of several travel modes, used four algorithms to train models based on different attributes as triaxial acceleration and resultant acceleration, and evaluated the recognition results referring to travel log.Subsequently, chose five instances each from three single and used two relative valid models to do recognition, the results pointed out that the recognition rate dropped in the order of walk-taxi-walk, walk-metro-walk and walk-bus-walk; and when used as the according attribute, triaxial acceleration performed better than resultant acceleration..At last chose a set of multiple travel mode data to train the model using decision tree algorithm, and applied the model directly to recognize another set of multiple travel mode data, and the result didn’t agree with the model. Then analyzed the result and used wavelet analysis to screen out travel segments, and apply decision tree algorithm to recognize each travel segment, then chose the mode of most recognized points among one segment as the mode of this one, and the recognition rate was over 80% no matter which attribute is applied; and when used as the according attribute, triaxial acceleration performed better than resultant acceleration.