基于深度学习的出行模式识别方法

Research on recognition method of transportation modes based on deep learning

摘要:

居民出行信息可体现居民活动规律、反映城市交通问题,是制定交通规划与管理的重要依据.利用GPS获取的轨迹数据虽具有大量时空信息但不能直接表达出行模式,需要数据处理和挖掘算法提取隐藏知识来识别出行模式.由于居民出行模式具有高度的非线性和复杂性,识别具有很大挑战.本文利用深度学习方法的特征学习表征优势,解决特征提取的繁琐计算或漏提特征等弊端,通过对轨迹进行去野和划分等预处理后,计算轨迹片段的运动学特征构成输入数据,提出基于卷积神经网络与门控循环单元相结合的识别出行模式方法,利用卷积神经网络的深层特征表征优势和门控循环单元的时序特性挖掘能力,提高对非线性分类问题的学习能力和识别出行模式的准确性.为验证所提出方法的有效性,还设计单独的卷积神经网络和门控循环单元等模型,在Geolife数据集上进行测试和对比.实验结果表明,本文方法虽仅计算4个特征量仍具有较好的识别效果,并且优于单独采用卷积神经网络等分类方法的识别性能.

Abstract:

Resident travel information can reflect the activity routines of residents and urban traffic problems, which is an important basis for formulating transportation planning and management. Although the trajectory information acquired by GPS has a lot of spatio-temporal information, it cannot directly express transportation modes. Data processing and mining algorithms are needed to extract hidden knowledge to infer transportation modes, while recognition has great challenges due to the high degree of non-linearity and complexity of residents’ travel patterns. In this study, the advantages of deep learning were utilized to solve difficult calculation features or missing extraction features. After pre-processing of the trajectory information, kinematic features of the trajectory segments were calculated to form the input data. A method that combines convolutional neural network with gate recurrent unit was proposed to recognize transportation modes. By utilizing the advantages of convolutional neural networks, the deep features and the ability of gate recurrent unit were characterized to mine time series characteristics, improve the learning ability of nonlinear classification problems, and increase the accuracy of transportation modes recognition. In order to verify the effectiveness of the proposed method, separate convolutional neural network and gate recurrent unit were designed, which was tested and compared on the published GeoLife dataset. Experimental results show that although the proposed method only used four features, it still received well recognition results. Besides, the proposed method had better recognition performance than using a convolutional neural network and other classification methods.