Traffic Travel Mode Identification Research Based On Multi-scale Feature Fusion

In the era of big data,the rapid development of computer technology changed the way people living and producing,meanwhile,it expanded the idea and the orientation of research.In terms of traffic governing,to solve widespread traffic problems such as traffic jam,traffic accident and terrible traffic environment,the scientific design and strategy should be utilized,which makes the research of citizens travel features the basis of itself.Collecting travel modes is part of gathering traveling information.Among researches of traveling modes recognition based on GPS data,the majority of them judge the travel modes by the features such as distance,traveling period,velocity and accelerate velocity,etc,which generalized as velocity-related features in this paper.This kind of velocityrelated features portrays the traveling situation of a certain traveling segment and classifies traveling modes refering to traveling situation features.To enrich the features using in classifying traveling modes and improve the accuracy of recognizing traveling modes,this paper introduces position sequence feature,the feature which contains the relationship among a points in a trajectory and its surrounding points.This kind of feature could enhance the discrimination of different traveling mdes.This paper propose the method based mixed features to recognize the travel modes.Firstly,put forward the extracting method of velocity-related features and location sequence features.The speedrelated features such as distance,time,speed,etc.They are calculated by GPS data.After that the features are reduced to 10 dimensionality using factor analysis.As for location sequence features,the 768 dimensionality location representation vector gained using the reference of BERT model,then utilizing Bi-LSTM + Attention network to integrate the location points vectors into 16 dimensionality location sequence features.The next step is merging two kinds of features and putting them in four different classify models which are SVM,BP neural network,Decision Tree and KNN.Finally BP Neural Network is the best model with accuracy 94.01% and 91.08% in training set and testing set respectively.Furthermore,compared with the model whose input is single velocity-related features,the merged feature models are always better than the single velocity-related features model.For example,the BP Neural Network using single velocity-related features with accuracy of 88.96% and 74.33% in training set and testing set respectively,which is lower than the merged feature as the input of BP Neural Network.To conclude,the method using mixed features to recognize travel modes is efficiency in experiments.The research result of this paper could be used in analyzing the rule of citizens traveling and referenced by scientific urban traffic governing.