EN中文

地球信息科学学报 >

2017 , Vol. 19 >Issue 3: 289 - 297

DOI: https://doi.org/10.3724/SP.J.1047.2017.00289

地球信息科学理论与方法

时空轨迹分类研究进展

  • 赵竹珺 ,
  • 吉根林 , *
展开
  • 南京师范大学计算机科学与技术学院,南京 210023
*通讯作者：吉根林（1964- ）,男,博士,教授,研究方向为数据挖掘及其应用。E-mail:

作者简介：赵竹珺（1993- ）,女,硕士生,研究方向为数据挖掘及其应用。E-mail:

收稿日期: 2016-08-09

  要求修回日期: 2016-12-05

  网络出版日期: 2017-03-20

基金资助

国家自然科学基金项目“云计算环境下顾及用户关系的手机用户时空轨迹模式挖掘方法研究”（41471371）

收起

Research Progress of Spatial-temporal Trajectory Classification

  • ZHAO Zhujun ,
  • JI Genlin , *
Expand
  • School of Computer Science and Technology, Nanjing Normal University, Nanjing 210023, China
*Corresponding author: JI Genlin, E-mail:

Received date: 2016-08-09

  Request revised date: 2016-12-05

  Online published: 2017-03-20

Copyright

《地球信息科学学报》编辑部 所有
Fold

摘要

时空轨迹分类旨在为一条轨迹预测类别。时空轨迹分类在城市规划、个性化用户推荐等方面具有重要应用价值,其过程主要包括轨迹数据预处理、特征提取、建立分类器3个阶段。本文综述了近年来时空轨迹分类的研究进展,首先对时空轨迹分类的过程进行概述;然后将时空轨迹分类算法按特征提取的方式分为基于运动特征的轨迹分类算法、基于分类规则的轨迹分类算法和基于图像信号分析的轨迹分类算法3类,分别论述了这些算法的基本思想和优缺点;之后对现有的轨迹分类算法从数据来源、分类器、特征提取方式等方面进行对比分析;最后讨论现有的时空轨迹分类算法面临的挑战。

关键词： 时空轨迹分类; 时空轨迹挖掘; 时空数据挖掘

本文引用格式

赵竹珺 , 吉根林 . 时空轨迹分类研究进展[J]. 地球信息科学学报, 2017 , 19(3) : 289 -297 . DOI: 10.3724/SP.J.1047.2017.00289

Abstract

Spatial-temporal trajectory classification aims at predicting the category of a spatial-temporal trajectory. The classification of spatial-temporal trajectories plays an important role in urban planning, personalized user recommendation and so on. The process of trajectory classification includes three stages: trajectory preprocessing, feature extraction and classification. This paper reviews the recent research progress on trajectory classification. Firstly, we introduce the process of trajectory classification. Then, the trajectory classification algorithms are classified into three categories according to the method of feature extraction, including the trajectory classification algorithm based on motion feature, the trajectory classification algorithm based on classification rule and the trajectory classification algorithm based on image signal analysis. We also discuss the basic ideas, advantages and disadvantages of these algorithms. Thirdly, we compare the existing classification algorithms according to the sensors, feature extraction and classifiers used in these algorithms. Finally, we introduce the challenges of the existing trajectory classification algorithms.

Key words： spatial-temporal trajectory classification; spatial-temporal trajectory mining; spatial-temporal mining

1 引言

近年来,随着智能采集设备的发展,研究人员很容易收集用户的时空轨迹数据,如可利用移动手机、全球定位系统GPS、加速度计等采集轨迹数据。轨迹数据挖掘的研究包括时空轨迹的聚类分析、伴随模式挖掘、频繁模式挖掘以及轨迹分类等。时空轨迹分类指通过训练轨迹数据建立一个模型,通过这个模型可以预测一条轨迹的类别。轨迹分类具有重要的应用价值,包括智能城市交通规划,识别用户轨迹的交通模式,如骑自行车或坐公交或小汽车[1-3],识别可疑车辆,识别飓风轨迹的强度值[4-5],识别出租车的载客、空载或停靠等状态[6],以及识别轨迹是否为异常轨迹等[7]。Krumm等[8]提出基于wifi数据的时空轨迹分类算法,使用隐马尔可夫模型将轨迹划分为停留与运动2类,Yin等[9]提出DBN模型,最低一层为wifi信号,基于该层向上挖掘分析,最终得到用户活动的目的。随着GPS精准定位的普及,大量的轨迹分类研究开始基于GPS数据,研究者们基于GPS数据提取轨迹的重要特征,建立分类模型,并得到了较高的准确率[10]。Gonzalez等[11]利用GPS收集数据,运用神经网络原理自动识别交通出行方式。其提出了一个关键点算法,过滤一些不必要的GPS数据,而只保留最重要的点的信息。之后,研究者使用平均速度、停留时间等特征以及关键点作为神经网络的输入,取得了较高的准确率[12]。如今,智能手机成为了人们的日常生活必备,而智能手机自带10多种传感器,使研究者可以获得多种数据,轨迹研究更加精确。 Assemi等[13]利用手机上的一个应用ATLAS II(Advanced Travel Logging Application for Smartphones II)收集数据,该应用运行于手机后台,记录手机的经度纬度、速度、方向、时间戳等信息,当用户停留时间超过一定阈值时停止记录。除了对传感器所收集到数据分析之外,Wang等[14]发现,大量的用户位置信息可以从CDR(Call Detail Records)-呼叫详细记录中收集到。该类数据在用户打电话、发短信或浏览网页时会自动产生,因此数据较便宜并且容易获取。通过该类数据研究者可以分析出用户的旅行信息-用户标识,目的地,开始时间,结束时间。
然而对时空轨迹的分类并不是一件容易的事,主要面临以下几个难点：
（1）首先时空轨迹数据本身的缺陷,如噪声数据、数据采样率不同,以及在一段时间内可能信号缺失。其次,在一些场合下,一条轨迹中存在多段不同类型的轨迹,如何对轨迹进行相应的预处理是需要研究的问题之一。
（2）时空轨迹分类需要具有标签的轨迹训练数据,而具有标签的轨迹训练数据量较少。特别是识别一条轨迹中包含的多种运动状态时,需要人为标识,想要获得较多的训练数据难度较大。如何建立合理的轨迹分类模型以应对轨迹标签数据较少的情况[15]。
（3）如何找到不同类型轨迹中最具有辨别力的特征,同时由于不同用户同种运动模式的轨迹特点可能不同,如何研究具有用户个体特点的轨迹分类问题[16]。
本文针对以上3个问题,综述时空轨迹的分类问题：首先介绍时空轨迹分类的主要过程;然后根据时空轨迹分类算法的特征提取方式出发,将轨迹分类算法分成3类,并分别介绍每种轨迹分类算法的特点;最后讨论现有方法面临的主要问题和挑战。

2 时空轨迹分类过程

时空轨迹分类的过程主要分为轨迹预处理、特征提取、建立分类器3个阶段（图1）。
Fig. 1 The process of trajectory classification

图1 轨迹分类过程图

轨迹预处理主要包括轨迹重采样、去噪、分段等过程。轨迹重采样指通过插值等手段使轨迹采样点间的时间间隔相同,便于后续轨迹挖掘。在一些使用多种数据综合考虑的场合,对数据进行重采样十分重要。轨迹去噪指去除轨迹数据中的噪声点,Lin等[17]使用Kalman Filter平滑轨迹,如图2所示,平滑后的走路段轨迹的速度累积分布函数与平滑前的累积分布函数对比图,发现平滑后走路明显更加符合人类的认知,走路段的速度均小于4 m/s。
Fig. 2 Cumulative distribution function of velocity[17]

图2 速度的累积分布函数图[17]

在一些时空轨迹分类任务中,一条轨迹可能存在多种类别,因此轨迹分段十分重要。Mountain等[2]认为一段轨迹在方向与速度上快速变化并在之后的轨迹段上速度与方向又趋于稳定时,代表移动模式的变化。所以分段算法首先计算轨迹中点与前一个点在速度与方向上的变化,如果大于一定阈值则为分段点。这种方法会产生较多分段点,如红路灯处或者交通堵塞的情况下,速度变化点会很多。
Zhu等[6]认为出租车的模式（载客或空车）转换必定需要停留点,所以通过识别停留点作为分段点。同时,Zhu考虑到红绿灯也会导致车辆停车,但并未发生任何模式转换,于是他提出使用监督学习识别停留点的方法,在识别过程中结合路网知识,为每个停留点计算5种特征值。
轨迹特征提取旨在提取最具有分类作用的轨迹特征。特征提取包括轨迹的运动特征、形状特征以及轨迹出现的位置特征、时间特征等。
建立分类器过程是将训练数据的特征向量作为分类器的输入,训练出分类模型。
Zheng等[18]使用决策树、支持向量机、贝叶斯网络、条件随机域4种分类手段分别将轨迹数据划分为走路、骑车、公交、私家车4种类型,结果发现决策树的分类效果最好。Jahangiri等[19]使用k近邻、支持向量机、决策树、Bagging与随机森林5种分类手段将轨迹划分为包括走路在内的5种类别,其准确率为90%以上。Shafique等[20]使用支持向量机、神经网络、决策树、Boosted Decision Tree、随机森林与朴素贝叶斯分别进行实验,发现随机森林与boosted decision trees分类精度最高,但是随机森林的计算速度更快,更适合轨迹分类。
除了使用单独的一种分类器完成轨迹分类任务以外,许多研究者使用多阶段分类的方式。 Reddy等[21]考虑到轨迹前后模式的转化概率,即一段轨迹属于某种类别的概率与前后段轨迹所属类别相关。首先使用决策树确定每一段属于某种运动模式的概率,然后将此概率作为隐马尔可夫模型的输入,实验最高精确度达到93.6%。

3 时空轨迹分类算法

时空轨迹分类的关键在于轨迹特征提取。本文按照特征提取的方法将时空轨迹分类算法分为3类：基于运动特征的轨迹分类算法、基于分类规则的轨迹分类算法、基于图像信号分析的轨迹分类算法。

3.1 基于运动特征的时空轨迹分类算法

本文的运动特征指轨迹的速度、角度和角速度等特征的平均数、最大数、中位数等可以反映运动对象运动特征的参数。
Zheng等[3,18]将一条轨迹划分为多种交通模式,因此在轨迹预处理阶段,需要首先对轨迹进行分段。他认为不同的移动模式转换间会存在一段走路段,因此以走路段的起点和终点作为分段点。首先通过计算每个采样点的速度区分走路点与非走路点,然后通过结合前后连续点找到走路段与非走路段。
在特征提取过程中,Zheng等发现不同的运动模式在停留的次数多少、运动方向变化程度等方面不同,从图3、4可看出,走路,开车与坐公交的停留次数存在较大差异;走路与开车的运动方向改变频率也存在较大差异。因此,提出了3种有效的特征值,分别为Heading change rate (HCR),Stop Rate (SR)与Velocity Change Rate (VCR),分别表示一段轨迹中运动方向超过一定阈值、速度小于一定阈值以及速度变化超过一定阈值的采样点个数的比例。
Fig. 3 Speed changes of different traffic modes[3]

图3 不同交通模式速度变化情况[3]

Fig. 4 Different trajectories of different traffic modes[3]

图4 不同交通模式轨迹[3]

Zheng等使用以上方法为每条轨迹段计算特征值,分别放入决策树、支持向量机、贝叶斯网络、条件随机域中进行实验。
Bolbol等[22]将轨迹划分为长度相同的子轨迹,根据方差分析(ANOVA)确定每一个特征变量的重要程度,得出速度与加速度是对分类具有重要意义的2个特征。该算法采用滑动窗口的方式对轨迹进行特征提取,最后使用SVM对轨迹进行分类。
朱进等[23]提出基于运动特征的轨迹分类方法。朱进将轨迹特征分为全局特征与局部特征2种：全局特征包括速度、加速度、曲率、方向和转角的均值、中位数、标准差,变异系数、最大的3个数、最小的3个数、自相关系数、偏度系数和峰度系数;局部特征先将轨迹分割成段并计算局部特征,其将运动参数表示为一个时间序列,反映了运动参数的幅度和频率随时间的变化。幅度表示运动参数的相对大小,而频率表示运动参数变化的快慢。根据幅度和频率的高低。时间序列上的点分为4类：高幅高频、高幅低频、低幅高频和低幅低频。对每条轨迹都分割成上述4类具有相似运动特性的轨迹段后,为每段计算统计量。
除了以上只使用了GPS数据提取特征外,Jahangiri等[19]在识别用户交通模式时不仅考虑了GPS数据,还考虑从智能手机上得到的加速度计数据、陀螺仪数据和旋转矢量3种数据。在使用多种数据的同时,需要对数据进行重采样,使多种数据在时间戳上一一对应。其一共提取了165种特征,包括每种数据在一段轨迹中的平均值、最大最小值、标准差、信号的强度等,并在实验阶段取得了较高的准确率与召回率。
Su等[16]提出了基于智能手机的交通模式识别算法。通过智能手机上的加速度计、重力传感器以及气压计得到特征值,并使用XSTND、YMAX、ZSTND 3个重要特征值,分别代表三维空间中的x、y、z轴上的加速度的标准差与最大值。
如图5所示,Su等将训练数据集分为2部分,即原始训练数据和该用户的训练数据。因为不同用户在不同的运动模式下的运动特征有所差异,如有些用户开车比较稳定,有些用户开车比较野性,轨迹中存在较大的加速度变化。其使用的识别模型如图5所示,使用所有用户的原始训练数据建立分类模型,每次增加一个该用户的训练数据,便更新一次分类模型。
Fig. 5 The classification model

图5 分类模型

除了结合多种传感器数据外,研究者将GPS数据与路网、POI兴趣点数据相结合,得到了较好的结果。
Zhu等[6]将GPS数据采样点投影到路网上,找到每个采样点所在的路段ID号。因此一段轨迹上路段ID的转移可以作为一个特征值。其次,结合兴趣区域数据（诸如星级酒店、停车点、超市以及娱乐场所等数据）,将采样点映射到其距离最近的兴趣区域。
Biljecki等[1]认为移动模式的转换中间应该存在停留点,同时信号缺失点也可能是移动模式转换点,因为模式的转化可能引起信号的短暂缺失,如在火车站里面信号不好。该分段算法计算每一个采样点的速度,如果连续12 s时间间隔内采样点的速度都未超过2 km/s,则认为该处发生了停留,标记为分段点。同时,查看连续采样点间的时间间隔,如果间隔大于30 s,则认为信号缺失,标记为分段点。但这种方法不能识别运动模式的转换中间没有经过短暂停留的情况。Biljecki利用GIS信息（开放街道地图数据,Open Street Map）来辅助GPS轨迹分类。该方法使用速度信息以及距交通网络（公共汽车线路、地铁线、道路路网）的邻近指数来作为轨迹的特征值。

3.2 基于分类规则的时空轨迹分类算法

在一些只考虑GPS数据的场合,为了研究每种类型轨迹的特征,研究者不只考虑轨迹的形状特征,还考虑到轨迹出现的位置特征。所谓分类规则,形如“模式–>类别”,指符合某种模式的轨迹大多数属于某种类别。通过分析训练数据,挖掘到支持度与置信度均大于一定阈值的分类规则[24],帮助建立特征向量。
该类方法通过考虑轨迹类别的模式挖掘方法发现对分类具有重要作用的分类规则,Fradkin等[25]、Cheng等[26-27]都提出了考虑类别的频繁模式挖掘算法。
找到分类规则之后,对每条轨迹依次查看是否符合每个分类规则,如果符合,则将该轨迹的特征向量的对应位置设置为相应规则的置信度,反之则设为零。
Brinkhoff[28]挖掘某种状态变化序列,满足这种状态变化序列的轨迹具有较大的概率属于某种类别。然后通过马尔可夫链确定轨迹类别。
Lee等[4]首先对轨迹进行考虑类别的分段,认为轨迹的一些重要属性可能存在于轨迹的部分段中,因此提出一种考虑轨迹标签的分段方法。如图6所示,图中虚线与实线表示不同的2类轨迹段。对于分类来说,明显前半段更加重要。因为与虚线同一类的轨迹不会出现在前半段,所以需要把实线分段,提取出前半段。该方法首先将轨迹段通过计算MDL（Minimum Description Length）值,找到在形状上发生较大变化的地方进行分段,然后考虑每一段的起点与终点附近的其他类型的段的个数是否大于一定阈值,如果大于阈值则分段。
Fig. 6 Trajectory segmentation for critical segments

图6 重要轨迹分段

通过对训练轨迹进行基于区域的聚集与基于轨迹的聚集,找到基于区域的分类规则以及基于轨迹运动模式的分类规则。基于区域的分类规则是指找到一些矩形区域,通过这些矩形区域的轨迹段大多属于一类。如图7所示,通过B、H区域的轨迹都是红色的,通过F区域的轨迹都是蓝色的。这些区域通过计算MDL值递归地划分矩形空间得到。基于轨迹的聚集是指将距离接近的轨迹段聚集,通过考虑轨迹段类别的DBSCAN算法得到几乎属于同一类型的轨迹簇。这种方法可以发现不同类型的轨迹段的重要特征,并且在动物轨迹识别以及飓风轨迹识别上取得了较高的准确率,但该方法并没有考虑时间信息。
Fig. 7 Regional classification rules[4]

图7 区域分类规则[4]

Lee等[29]认为连续的路网序列模式对轨迹分类具有较大作用。连续的路网序列模式指一个具有一定顺序的路段序列,且通过这个序列的训练轨迹个数达到一定数量。对于找到的这些模式,通过对比符合每个模式的正例与反例计算F-score,筛选出有较强识别能力的特征。该方法适用于二分类问题,而多分类问题的效率高低并不明确。
Patel[5]在Lee等[4]提出的方法基础上进行改进,认为在挖掘频繁模式时加入持续时间可以帮助我们区分具有不同速度的运动对象。该方法通过对训练轨迹进行考虑持续时间的区域分类规则挖掘与基于持续时间的路径分类规则发现,对轨迹提取特征。基于考虑持续时间的区域分类规则指找到一些矩形区域,通过该区域且通过该区域的时长在规定范围之内的轨迹大部分属于同一类型。基于持续时间的路径分类规则发现指找到一些路径,如从A点集经过t时间到B点集,符合这个模式的轨迹大部分属于一类。
Patel首先通过划分矩形区域的方法找到置信度大于一定阈值的区域分类规则,然后通过对训练轨迹集建立轨迹网络图,找到路径分类规则。结合找到的分类规则以及轨迹的全局特征（平均速度,持续时间等）对每条轨迹形成特征向量,放入分类器中。该算法的流程图如图8所示。
Fig. 8 The flow diagram of Patel D algorithm

图8 Patel D算法流程图

3.3 基于图像信号分析的时空轨迹分类算法

该类轨迹分类算法将轨迹路径看作信号序列或者图像形式,并试图确定产生信号或图像的轨迹类别。首先将轨迹信息转换为信号或图像,并从中提取轨迹的特征。该类轨迹分类手段面临的主要问题是：① 如何将轨迹转换为图像或信号;② 如何处理不同长度的轨迹;③ 如何提取特征,对于每一段轨迹应该提取何种特征;④ 如何将时间信息考虑在轨迹分类当中。
这种类型的轨迹分类算法一般更注重轨迹的几何形状信息,忽略时间与地点信息。
Macdonald等[30]首次将图像处理上常用到的矩与小波应用到轨迹分类上。首先将轨迹路径转化为信号,并设定一个步长20,将轨迹分隔成子轨迹。对每一段子轨迹提取包括回归系数在内的5种不同特征,最后对于每一条轨迹,取它的每一个子轨迹段的5种特征值的平均值与标准差作为该条轨迹的特征值。这种方法可以应对不同长度的轨迹并且不会缺失重要的长度信息。
Endo等[31]提出使用深度神经网络（DNN）自动提取特征的方法。该方法首先将轨迹转换为图像形式。图像被分为网格,每个网格的像元大小表示为轨迹在该网格中停留的时间长短。然后将图像转换为矩阵,通过DNN得到特征值。具体过程如图9所示。
Fig. 9 The process of changing a trajectory into an image[31]

图9 轨迹转换为图像过程[31]

4 时空轨迹分类算法比较

3种时空轨迹分类算法都有各自的优点与局限性。基于运动特征的轨迹分类算法需要提取轨迹中具有代表性的运动特征,这些特征对分类有重要作用,但它无法解决在一些特殊情况下不同类型的轨迹运动特征相似的情况,如堵车情况下汽车运动特征类似于走路。
基于分类规则的轨迹分类算法需要利用轨迹出现的位置以及时间信息,因此可发现停车场等重要地点,同时也可以区分不同速度的运动对象。
基于图像信号分析的轨迹分类算法更注重轨迹的形状特征,只考虑采样点出现的时间顺序信息,不需要轨迹出现的地点以及具体的时间信息。
本文从轨迹数据来源（传感器）、特征提取、分类器和每种分类方法的优缺点4方面对时空轨迹分类算法进行比较（表1）。从表1可看出,GPS收集到的轨迹数据仍为当前轨迹分类的主流数据。
Tab. 1 Comparison of classification algorithms of different trajectories

表1 各轨迹分类算法比较

算法类别 轨迹采集方式 分类器 轨迹特征 作者 优缺点
运动特征 GPS 专家系统 最大速度的95%,平均速度,平均移动速度,段出现的位置是否有基础设施且接近程度值 Biljecki F[1] 该类方法结合轨迹的特点以及统计学知识,提取具有较强识别能力的运动特征,但没有考虑到轨迹出现的地点以及具体时间信息
GPS 决策树、支持向量机、贝叶斯网络、条件随机域 角度变化率 (HCR),停留次数比例 (SR)与速度变化率 (VCR) Zheng Y[3, 18]
GPS 无监督学习 最大速度分布情况 Lin M[17]
GPS 决策树+隐马尔可夫 单条轨迹特征,前后段轨迹类别转换特征,地理数据特征 Zhu Y[6]
GPS、加速度计数据、陀螺仪数据 k近邻、支持向量机、决策树、Bagging与随机森林 加速度、速度、谱熵等165个特征值 Jahangiri A[19]
GPS 支持向量机 运动轨迹的全局特征与局部特征 朱进[23]
GPS 支持向量机 加速度、减速度、加速度和减速度标准偏差 Sun Z[32]
加速度计、 陀螺仪 支持向量机、神经网络、决策树、Boosted Decision Tree、随机森林与朴素贝叶斯 两传感器数据标准差、偏度和峰度 Shafique M A[20]
GPS 神经网络 平均速度、停留时间、关键点信息等 Gonzalez P A[11]
GPS 高斯混合模型 速度信息 Zhu W[15]
GPS,加速度计 决策树+隐马尔可夫模式 均值,方差等 Reddy S[21]
GPS、真实交通数据 贝叶斯、决策树、随机森林 速度信息、平均车辆亲密度、平均铁路亲密度、平均候选车辆亲密度 Stenneth L[33]
GPS 支持向量机 速度、加速度 Bolbol A[22]
gms、wifi 决策树 基站信息、信号强度、接入wifi的持续时间等 Mun M[34]
分类规则 GPS、呼叫详细记录 支持向量机 起点位置、终点位置、从起点到终点的运动时间 Wang H[14] 该类方法充分利用训练数据,挖掘具有较强识别能力的分类规则,考虑到轨迹的时间以及地点信息,但仅适用于训练数据足够充分,便于挖掘规律的情况
GPS 支持向量机 区域分类规则、路径分类规则 Brinkhoff T[28]、Lee J G[4]、
Patel D[5]
图像信号 加速度计 Adaboost+隐马尔可夫 基于帧的特征、基于峰值的特征(加速度)、基于段的特征值 Hemminki S[35] 该类方法更加注重轨迹形状对轨迹分类结果的影响,但轨迹转化为图像或信号时,并不能将时间信息考虑进去
GPS 支持向量机 墨西哥帽小波、回归系数等5种特征的平均值与标准差 Macdonald A[30]
GPS 逻辑回归、支持向量机、决策树 转化为图像,使用神经网络自动生成特征 Endo Y[31]
大多数轨迹分类算法属于基于运动特征的轨迹分类算法,因为该种方法考虑到轨迹的特殊性,提取轨迹的运动特征,具有较好的识别能力。对于不同传感器采集的轨迹数据,提取的轨迹特征可能不同,其轨迹分类的结果也可以不同。基于分类规则的轨迹分类算法需要轨迹数据训练集比较大。基于图像信号分析的轨迹分类算法需要通过合理的方法将轨迹转换为图像或信号形式。
此外,使用支持向量机作为分类器的轨迹分类算法较多。针对一条轨迹中可能存在多种类别的轨迹段时,有些算法使用多阶段分类的方法,先使用一种分类器得到轨迹属于每一种类别的概率,再通过隐马尔可夫模型,依据轨迹前后段的类别计算出轨迹类别的最终概率。

5 结论与展望

时空轨迹数据的采集方式从使用WIFI到利用智能手机上多种传感器进行采集,使采集到的轨迹数据越来越精确,从而提高了轨迹分类的准确度。但由于带标签的轨迹数据收集的困难,半监督的时空轨迹分类方法研究成为当前的研究热点。
时空轨迹分类总体分为数据预处理、特征提取、分类器建立3个阶段。根据特征提取方法不同,轨迹分类算法可以分为基于运动特征的轨迹分类算法、基于分类规则的轨迹分类算法和基于图像信号分析的轨迹分类算法3种。而这3种方法都有各自的优点与局限性,都存在一定的改进空间,如在挖掘轨迹分类规则以及将轨迹转化为图像信号时,需要考虑时间信息。
对轨迹分类的研究仍面临以下挑战：
（1）如何综合利用多种传感器采集的时空轨迹数据来提高轨迹分类的精度。对于带标签的轨迹数据收集困难的情况,如何利用半监督学习方法建立时空轨迹分类器。
（2）如何提取具有较强识别能力的轨迹特征,使这些轨迹特征能够应对环境变化而产生的特殊情况,如堵车情况下汽车轨迹、自行车轨迹、人步行轨迹三者相似的情况等。
（3）在线时空轨迹分类算法的研究[36]。对于流式轨迹数据,存在一条轨迹具有多种类别轨迹段的情况,需要研究如何采用滑动窗口算法对轨迹进行合理分段;如何设计增量式学习算法建立轨迹分类器,达到较高的识别精度和效率。

The authors have declared that no competing interests exist.

参考文献

原文顺序 | 文献年度倒序 | 文中引用次数倒序
[1]
Biljecki F, Ledoux H, Van Oosterom P.Transportation mode-base segmentation and classification of movement trajectories[J]. International Journal of Geographical Information Science, 2013,27(2):385-407.The knowledge of the transportation mode used by humans (e. g. bicycle, on foot, car, and train) is critical for travel behaviour research, transport planning and traffic management. Nowadays, new technologies such as the GPS have replaced traditional survey methods (paper diaries, telephone) since they are more accurate and problems such as underreporting are avoided. However, although the movement data collected (timestamped positions in digital form) have generally high accuracy, they do not contain the transportation mode. We present in this paper a new method for segmenting movement data into single-mode segments and to classify them according to the transportation mode used. Our fully automatic method differs from previous attempts for five reasons: (1) it relies on fuzzy concepts found in expert systems, i. e. membership functions and certainty factors; (2) it uses OpenStreetMap data to help the segmentation and classification process; (3) we can distinguish between 10 transportation modes (incl. between tram, bus, and car) and we propose a hierarchy; (4) it handles data with signal shortages and noise, and other real-life situations; (5) in our implementation, there is a separation between the reasoning and the knowledge, so that users can easily modify the parameters used and add new transportation modes. We have implemented the method and tested it with a 17-million point dataset collected in the Netherlands and elsewhere in Europe. The accuracy of the classification with the developed prototype, determined with the comparison of the classified results with the reference data derived from manual classification, is 91.6 percent.

DOI

[2]
Mountain D, Raper J.Modelling human spatio-temporal behaviour: A challenge for location-based services[C]//Proceedings of 6th International Conference on Geocomputation. Brisbane: Australia, 2001:24-26.

[3]
Zheng Y, Chen Y, Li Q, et al.Understanding transportation modes based on GPS data for web applications[J]. Acm Transactions on the Web, 2010,4(1):495-507.User mobility has given rise to a variety of Web applications, in which the global positioning system (GPS) plays many important roles in bridging between these applications and end users. As a kind of human behavior, transportation modes, such as walking and driving, can provide pervasive computing systems with more contextual information and enrich a user's mobility with informative knowledge. In this article, we report on an approach based on supervised learning to automatically infer users' transportation modes, including driving, walking, taking a bus and riding a bike, from raw GPS logs. Our approach consists of three parts: a change point-based segmentation method, an inference model and a graph-based post-processing algorithm. First, we propose a change point-based segmentation method to partition each GPS trajectory into separate segments of different transportation modes. Second, from each segment, we identify a set of sophisticated features, which are not affected by differing traffic conditions (e.g., a person's direction when in a car is constrained more by the road than any change in traffic conditions). Later, these features are fed to a generative inference model to classify the segments of different modes. Third, we conduct graph-based postprocessing to further improve the inference performance. This postprocessing algorithm considers both the commonsense constraints of the real world and typical user behaviors based on locations in a probabilistic manner. The advantages of our method over the related works include three aspects. (1) Our approach can effectively segment trajectories containing multiple transportation modes. (2) Our work mined the location constraints from user-generated GPS logs, while being independent of additional sensor data and map information like road networks and bus stops. (3) The model learned from the dataset of some users can be applied to infer GPS data from others. Using the GPS logs collected by 65 people over a period of 10 months, we evaluated our approach via a set of experiments. As a result, based on the change-point-based segmentation method and Decision Tree-based inference model, we achieved prediction accuracy greater than 71 percent. Further, using the graph-based post-processing algorithm, the performance attained a 4-percent enhancement.

DOI

[4]
Lee J G, Han J, Li X, et al.TraClass: trajectory classification using hierarchical region-based and trajectory-based clustering[J]. Proceedings of the Vldb Endowment, 2008,1(1):1081-1094.Trajectory classification, i.e., model construction for predict-ing the class labels of moving objects based on their trajecto-ries and other features, has many important, real-world ap-plications. A number of methods have been reported in the literature, but due to using the shapes of whole trajectories for classification, they have limited classification capability when discriminative features appear at parts of trajectories or are not relevant to the shapes of trajectories. These situ-ations are often observed in long trajectories spreading over large geographic areas. Since an essential task for effective classification is gen-erating discriminative features, a feature generation frame-work TraClass for trajectory data is proposed in this pa-per, which generates a hierarchy of features by partition-ing trajectories and exploring two types of clustering: (1) region-based and (2) trajectory-based. The former captures the higher-level region-based features without using move-ment patterns, whereas the latter captures the lower-level trajectory-based features using movement patterns. The proposed framework overcomes the limitations of the previ-ous studies because trajectory partitioning makes discrimi-native parts of trajectories identifiable, and the two types of clustering collaborate to find features of both regions and sub-trajectories. Experimental results demonstrate that TraClass generates high-quality features and achieves high classification accuracy from real trajectory data.

DOI

[5]
Patel D.Incorporating duration and region association information in trajectory classification[J]. Journal of Location Based Services, 2012,7(4):246-271.Trajectory classification has many useful applications. Existing works on trajectory classification do not consider the duration information of trajectory. In this paper, we extract duration-aware features from trajectories to build a classifier. Our method utilizes information theory to obtain regions where the trajectories have similar speeds and directions. Further, trajectories are summarized into a network based on the MDL principle that takes into account the duration difference among trajectories of different classes. A graph traversal is performed on this trajectory network to obtain the top-k covering path rules for each trajectory. Based on the discovered regions and top-k path rules, we build a classifier to predict the class labels of new trajectories. Experiment results on real-world datasets show that the proposed duration-aware classifier can obtain higher classification accuracy than the state-of-the-art trajectory classifier.

DOI

[6]
Zhu Y, et al.Inferring taxi status using GPS trajectories. Tech.Rep.MSR-TR-2011-144, Microsoft Research, Asia, March 2011.

[7]
Lee J G, Han J, Li X. Trajectory outlier detection: Apartition and detect framework[C]//In: Proceedings of the 24th International Conference on Data Engineering. Washington:USA, 2008:140-149.

[8]
Krumm J, Horvitz E.LOCADIO: Inferring motion and location from Wi-Fi signal strengths[C]// International Conference on Mobile and Ubiquitous Systems: NETWORKING and Services. Boston: USA, 2004:4-13.

[9]
Yin J, Chai X, Yang Q.High-level goal recognition in a wireless LAN[C]// Nineteenth National Conference on Artificial Intelligence, Sixteenth Conference on Innovative Applications of Artificial Intelligence, 2004:578-584.

[10]
Zheng Y.Trajectory data mining: An overview[J]. Acm Transactions on Intelligent Systems & Technology, 2015,6(3):1-41.The advances in location-acquisition and mobile computing techniques have generated massive spatial trajectory data, which represent the mobility of a diversity of moving objects, such as people, vehicles, and animals. Many techniques have been proposed for processing, managing, and mining trajectory data in the past decade, fostering a broad range of applications. In this article, we conduct a systematic survey on the major research into trajectory data mining, providing a panorama of the field as well as the scope of its research topics. Following a road map from the derivation of trajectory data, to trajectory data preprocessing, to trajectory data management, and to a variety of mining tasks (such as trajectory pattern mining, outlier detection, and trajectory classification), the survey explores the connections, correlations, and differences among these existing techniques. This survey also introduces the methods that transform trajectories into other data formats, such as graphs, matrices, and tensors, to which more data mining and machine learning techniques can be applied. Finally, some public trajectory datasets are presented. This survey can help shape the field of trajectory data mining, providing a quick understanding of this field to the community.

DOI

[11]
Gonzalez P A, Weinstein J S, Barbeau S J, et al.Automating mode detection using neural networks and assisted GPS data collected using GPS-enabled mobile phones[C]// 15th World Congress on Intelligent Transport Systems and ITS America's 2008 Annual Meeting. New York, USA, 2008:30267-30279

[12]
肖艳丽,张振宇,杨文忠.移动数据的交通出行方式识别方法[J].智能系统学报,2014,9(5):536-543.识别用户出行的交通方式,对理解用户移动性、交通状况的分析和预测、社会活动模式挖掘等方面起着非常关键的作用。随着无线网络技术的快速发展,越来越多的传感器被用于收集移动数据,如何通过收集的信息准确地识别用户不同的交通出行方式,近年来得到了广泛的研究。针对已有的从不同角度识别交通方式的方法,首先介绍了每种方法的具体内容及应用,然后对不同方法进行分类研究,并重点分析了各类方法的特点,分析几种不同方法在不同条件下的识别精确度,最后,给出了交通方式识别方法的进一步研究方向。

DOI

[ Xiao Y L, Zhang Z Y, Yang W Z.Research of the identification methods for transportation modes based on mobile data[J]. CAAI Transactions on Intelligent Systems, 2014,9(5):536-543. ]

[13]
Assemi B, Safi H, Mesbah M, et al.Developing and validating a statistical model for travel mode identification on smartphones[J]. IEEE Transactions on Intelligent Transportation Systems, 2016,17(7):1920-1931.Smartphone travel surveys are able to capture accurate details about individuals' travel behavior. However, extracting the required information (e.g., travel mode and purpose) from the data captured by smartphone applications is relatively complex, particularly when relying on the computational power of smartphones and limiting the communications between these applications and third parties [e.g., geographic information systems (GIS)]. These limitations are mainly enforced to enable passive data collection through smartphones by automatically recognizing the mode and purpose of trips. Furthermore, limited data transfer between the application and third parties ensures the privacy protection of survey participants and facilitates real-world travel surveys with large sample sizes. Accordingly, the objective of this paper is to develop a model of travel mode identification, which can be integrated with smartphone travel surveys without using GIS data or interacting with participants. Most existing models and algorithms are either inaccurate or computationally complex, and require extensive processing power. A smartphone travel survey, namely, the Advanced Travel Logging Application for Smartphones II (ATLAS II), has been used to collect individuals' travel data across New Zealand and Queensland, Australia. A detailed algorithm is put forward to clean the captured data, segment trips into single modal trips, and develop multiple statistical models for comparison, using the data collected from New Zealand. The preferred approach, which is adapted for the integration with smartphone travel survey applications, is validated using the two separate data sets from New Zealand and Australia. The resulting mode identification model (i.e., a nested logit model with eight variables) could detect travel modes with the accuracy of 97% for New Zealand after preprocessing (i.e., data cleaning and trip segmentation) and 79.3% for Australia without any preprocessing.

DOI

[14]
Wang H, Calabrese F, Di Lorenzo G, et al.Transportation mode inference from anonymized and aggregated mobile phone call detail records[C]//13th International IEEE Conference on Intelligent Transportation Systems (ITSC),Funchal: Portugal, 2010:318-323.

[15]
Zhu W, Ash J, Li Z, et al.Applying semi-supervised learning method for cellphone-based travel mode classification[C]//Smart Cities Conference (ISC2).London:UK, 2015:1-6.

[16]
Su X, Caceres H, Tong H, et al.Online travel mode identification using smartphones with battery saving considerations[J]. IEEE Transactions on Intelligent Transportation Systems, 2016,99(1):1-14.Personal trips in modern urban society usually involve multiple travel modes. Recognizing a traveler's transportation mode is not only critical to personal context awareness in related applications but also essential to urban traffic operations, transportation planning, and facility design. While most current practice often leverages infrastructure-based fixed sensors or a Global Positioning System (GPS) for traffic mode recognition, the emergence of the smartphone provides an alternative promising way with its ever-growing computing, networking, and sensing power. In this paper, we propose a GPS-and-network-free method to detect a traveler's travel mode using mobile phone sensors. Our application is built on the latest Android platform with multimodality sensors. By developing a hierarchical classification method with an online learning model, we achieve almost 100% accuracy in a binary classification of wheeled/unwheeled travel modes and an average of 97.1% accuracy with all six travel modes (buses, subways, cars, bicycling, walking, and jogging). Our system (a) could adapt to each traveler's pattern by using the online learning model, and it performs significantly faster in computation than the offline model, and (b) works with a low sampling frequency for sensing so that it saves the smartphone battery.

DOI

[17]
Lin M, Hsu W J, Lee Z Q.Detecting modes of transport from unlabeled positioning sensor data[J]. Journal of Location Based Services, 2013,7(4):272-290.Global positioning systems (GPS) logs recorded in personal devices contain rich information such as travel patterns, locations of frequent visits and place–event associations. There have been rather successful attempts in detecting the mode of transport from GPS logs such as walking, driving or taking a bus, which has found varied applications. However, the best-known schemes either require tedious manual labelling or pre-training process (or both). We present MoDetect (MD), a unsupervised scheme which eliminates the need of manual labelling and pre-training while attaining equal or greater accuracy compared with the best-known supervised methods. MD can also cater for differences in individual's behaviours, and hence may be more widely applicable than the existing schemes. To achieve this, MD relies on Kolmogorov–Smirnov test which offers a theoretical assurance when computing similarity between segments of records. Our analysis shows that the higher speed modes can be better differentiated through a weighted bootstrapping procedure. We also augment the decisions with reference to the transfer probabilities between different modes at locations identified from the GPS records.

DOI

[18]
Zheng Y, Li Q, Chen Y, et al.Understanding mobility based on GPS data[C]//Ubiquitous Computing, International Conference, UBICOMP, Seoul: Korea, 2008:312-321.

[19]
Jahangiri A, Rakha H A.Applying machine learning techniques to transportation mode recognition using mobile phone sensor data[J]. IEEE transactions on intelligent transportation systems, 2015,16(5):2406-2417.This paper adopts different supervised learning methods from the field of machine learning to develop multiclass classifiers that identify the transportation mode, including driving a car, riding a bicycle, riding a bus, walking, and running. Methods that were considered include K-nearest neighbor, support vector machines (SVMs), and tree-based models that comprise a single decision tree, bagging, and random forest (RF) methods. For training and validating purposes, data were obtained from smartphone sensors, including accelerometer, gyroscope, and rotation vector sensors. K-fold cross-validation as well as out-of-bag error was used for model selection and validation purposes. Several features were created from which a subset was identified through the minimum redundancy maximum relevance method. Data obtained from the smartphone sensors were found to provide important information to distinguish between transportation modes. The performance of different methods was evaluated and compared. The RF and SVM methods were found to produce the best performance. Furthermore, an effort was made to develop a new additional feature that entails creating a combination of other features by adopting a simulated annealing algorithm and a random forest method.

DOI

[20]
Shafique M A, Hato E.A Comparison among various classification algorithms for travel mode detection using sensors' data collected by smartphones[C]//International Conference on Computers in Urban Planning and Urban Management, 2015.

[21]
Reddy S, Mun M, Burke J, et al.Using mobile phones to determine transportation modes[J]. Acm Transactions on Sensor Networks, 2010,6(2):662-701.As mobile phones advance in functionality and capability, they are being used for more than just communication. Increasingly, these devices are being employed as instruments for introspection into habits and situations of individuals and communities. Many of the applications enabled by this new use of mobile phones rely on contextual information. The focus of this work is on one dimension of context, the transportation mode of an individual when outside. We create a convenient (no specific position and orientation setting) classification system that uses a mobile phone with a built-in GPS receiver and an accelerometer. The transportation modes identified include whether an individual is stationary, walking, running, biking, or in motorized transport. The overall classification system consists of a decision tree followed by a first-order discrete Hidden Markov Model and achieves an accuracy level of 93.6% when tested on a dataset obtained from sixteen individuals.

DOI

[22]
Bolbol A, Cheng T, Tsapakis I, et al.Inferring hybrid transportation modes from sparse GPS data using a moving window SVM classification[J]. Computers Environment & Urban Systems, 2012,36(6):526-537.Understanding travel behaviour and travel demand is of constant importance to transportation communities and agencies in every country. Nowadays, attempts have been made to automatically infer transportation modes from positional data, such as the data collected by using GPS devices so that the cost in time and budget of conventional travel diary survey could be significantly reduced. Some limitations, however, exist in the literature, in aspects of data collection (sample size selected, duration of study, granularity of data), selection of variables (or combination of variables), and method of inference (the number of transportation modes to be used in the learning). This paper therefore, attempts to fully understand these aspects in the process of inference. We aim to solve a classification problem of GPS data into different transportation modes ( car , walk , cycle , underground , train and bus ). We first study the variables that could contribute positively to this classification, and statistically quantify their discriminatory power. We then introduce a novel approach to carry out this inference using a framework based on Support Vector Machines (SVMs) classification. The framework was tested using coarse-grained GPS data, which has been avoided in previous studies, achieving a promising accuracy of 88% with a Kappa statistic reflecting almost perfect agreement.

DOI

[23]
朱进,江南,胡斌.移动对象多种运动参数在轨迹分类的应用[J].地球信息科学学报,2016,18(2):143-150.<p>轨迹分类是以训练轨迹的特征来预测未知轨迹的类标签,可进行可疑车辆识别、非法渔船检测和交通模式检测等重要应用。当前大多数轨迹分类方法只考虑速度和加速度这2个运动参数,且只利用简单的统计量（如均值、中值、最大值）,不能充分挖掘轨迹的潜在特点,导致分类精度不高。针对该问题,本文在总结相关文献的基础上,提出一种基于移动对象运动特征的轨迹分类方法。针对速度、加速度、曲率、方向和转角这5个运动参数,利用偏度系数、峰度系数、变异系数和时间序列分析其中的自相关系数等,提取有区分力的全局运动特征;并从分割后的子轨迹中提取局部运动特征。对于方向和转角,引入方向统计学对其运动特征进行精确计算。实验表明本方法在船舶、野生动物和飓风数据集的分类精度达到了100%、80%和71.43%,实验验证了本方法构建的运动特征,在不同数据集下有效可行。</p>

DOI

[ Zhu J, Jiang N, Hu B.The application of multiple movement parameters in trajectory classification for moving objects[J]. Journal of Geo-information Science, 2016,18(2):143-150. ]

[24]
Lin M, Hsu W J.Mining GPS data for mobility patterns: A survey[J]. Pervasive & Mobile Computing, 2014,12(11):1-16.With the help of various positioning tools, individuals’ mobility behaviors are being continuously captured from mobile phones, wireless networking devices and GPS appliances. These mobility data serve as an important foundation for understanding individuals’ mobility behaviors. For instance, recent studies show that, despite the dissimilarity in the mobility areas covered by individuals, there is high regularity in the human mobility behaviors, suggesting that most individuals follow a simple and reproducible pattern. This survey paper reviews relevant results on uncovering mobility patterns from GPS datasets. Specially, it covers the results about inferring locations of significance for prediction of future moves, detecting modes of transport, mining trajectory patterns and recognizing location-based activities. The survey provides a general perspective for studies on the issues of individuals’ mobility by reviewing the methods and algorithms in detail and comparing the existing results on the same issues. Several new and emergent issues concerning individuals’ mobility are proposed for further research.

DOI

[25]
Fradkin D, Mörchen F.Mining sequential patterns for classification[J]. Knowledge & Information Systems, 2015,45(3):731-749.While a number of efficient sequential pattern mining algorithms were developed over the years, they can still take a long time and produce a huge number of patterns, many of which are redundant. These properties are especially frustrating when the goal of pattern mining is to find patterns for use as features in classification problems. In this paper, we describe BIDE-Discriminative, a modification of BIDE that uses class information for direct mining of predictive sequential patterns. We then perform an extensive evaluation on nine real-life datasets of the different ways in which the basic BIDE-Discriminative can be used in real multi-class classification problems, including 1-versus-rest and model-based search tree approaches. The results of our experiments show that 1-versus-rest provides an efficient solution with good classification performance.

DOI

[26]
Cheng H, Yan X, Han J, et al.Discriminative frequent pattern analysis for effective classification[C]//IEEE, International Conference on Data Engineering. Istanbul:Turkey, 2007:716-725.

[27]
Cheng H, Yan X, Han J, et al.Direct discriminative pattern mining for effective classification[C]//2008 IEEE 24th International Conference on Data Engineering. Cancun: Mexico, 2008:169-178.

[28]
Brinkhoff T.A framework for generating network-based moving objects[J]. Geoinformatica, 2002,6(2):153-180.<a name="Abs1"></a>Benchmarking spatiotemporal database systems requires the definition of suitable datasets simulating the typical behavior of moving objects. Previous approaches for generating spatiotemporal data do not consider that moving objects often follow a given network. Therefore, benchmarks require datasets consisting of such <img src="/content/L82B200GX12CHQCE/xxlarge8220.gif" alt="ldquo" align="MIDDLE" border="0">network-based<img src="/content/L82B200GX12CHQCE/xxlarge8221.gif" alt="rdquo" align="MIDDLE" border="0"> moving objects. In this paper, the most important properties of network-based moving objects are presented and discussed. Essential aspects are the maximum speed and the maximum capacity of connections, the influence of other moving objects on the speed and the route of an object, the adequate determination of the start and destination of an object, the influence of external events, and time-scheduled traffic. These characteristics are the basis for the specification and development of a new generator for spatiotemporal data. This generator combines real data (the network) with user-defined properties of the resulting dataset. A framework is proposed where the user can control the behavior of the generator by re-defining the functionality of selected object classes. An experimental performance investigation demonstrates that the chosen approach is suitable for generating large data sets.

DOI

[29]
Lee J G, Han J, Li X, et al.Mining discriminative patterns for classifying trajectories on road networks[J]. IEEE Transactions on Knowledge & Data Engineering, 2011,23(5):713-726.

[30]
Macdonald A, Ellen J.Multi-level resolution features for classification of transportation trajectories[C]//IEEE, International Conference on Machine Learning and Applications. Miami: FL, 2015:713-718.

[31]
Endo Y, Toda H, Nishida K, et al.Deep feature extraction from trajectories for transportation mode estimation[C]//Pacific-Asia Conference on Knowledge Discovery and Data Mining. Auckland, New:Zealand, 2016:54-66.

[32]
Sun Z, Ban X.Vehicle classification using GPS data[J]. Transportation Research Part C Emerging Technologies, 2013,37(3):102-117.Vehicle classification information is crucial to transportation planning, facility design, and operations. Traditional vehicle classification methods are either too expensive to be deployed for large areas or subject to errors under specific situations. In this paper, we propose methods to classify vehicles using GPS data extracted from mobile traffic sensors, which is considered to be low-cost especially for large areas of urban arterials. It is found that features related to the variations of accelerations and decelerations (e.g., the proportions of accelerations and decelerations larger than 1 meter per square second, and the standard deviations of accelerations and decelerations) are the most effective in terms of vehicle classification using GPS data. By classifying general trucks from passenger cars, the average misclassification rate is about 1.6% for the training data, and 4.2% for the testing data.

DOI

[33]
Stenneth L, Wolfson O, Yu P S, et al.Transportation mode detection using mobile phones and GIS information[C]//ACM Sigspatial International Symposium on Advances in Geographic Information Systems, Chicago: USA, 2011:54-63.

[34]
Mun M Y, Seo Y W.Everyday mobility context classification using radio beacons[C]//Consumer Communications and NETWORKING Conference. IEEE, Las Vegas: NV, 2013:112-117.

[35]
Hemminki S, Nurmi P, Tarkoma S.Accelerometer-based transportation mode detection on smartphones[C]//ACM Conference on Embedded Networked Sensor Systems.Rome: Italy, 2013:1-14.

[36]
Boukhechba M, Bouzouane A, Bouchard B, et al.Online recognition of people's activities from raw GPS data: Semantic Trajectory Data Analysis[C]//Proceedings of the 8th ACM International Conference on Pervasive Technologies Related to Assistive Environments, 2015:575-578.

Options
文章导航

/

〈 〉