轨迹预测与意图挖掘问题研究

发布时间：2018-01-04 17:25

本文关键词：轨迹预测与意图挖掘问题研究　出处：《山东大学》2016年博士论文　论文类型：学位论文

【摘要】：随着定位设备和视频捕捉设备的广泛应用,时空轨迹数据的获得变得越来越便捷。例如,人们在社交网站(如Foursquare)上和朋友们分享自己的位置,这就产生了大量的签到数据。又如,车辆会被城市道路上的交通监控摄像机所捕捉,随后利用光学字符识别技术被处理成结构化的记录。签到数据和车辆通行数据都是典型的时空轨迹数据,他们都包含对象、位置以及时间三种属性。这些数据蕴含着巨大的价值,在城市计算、路径规划、位置预测等领域出现了大量的成果。本文针对时空轨迹挖掘领域的相关问题进行了深入的研究。首先,基于用户的历史轨迹研究下一个位置预测问题。在实际应用中,提前知晓用户的下一个位置,可以给用户推荐更合理的驾驶路线以及给用户推荐将要到达位置的广告信息。为了进行下一个位置预测,首先提出了全局马尔科夫模型(GMM)和个人马尔科夫模型(PMM)。GMM利用用户所有的轨迹发现集体移动模式；PMM利用每个用户自己的轨迹发现个人移动模式。这两个模型通过线性回归组合起来产生最后的预测器(NLPMM)。此外,由于时间因素对用户的移动模式有影响,我们寻求在模型中加入时间因素以提高预测的准确性。我们提出了三种方法来聚类不同时间段中的轨迹,并训练了更加细化的模型来预测下一个位置。进一步地,本文提出了融合用户相似度和轨迹相似度的马尔科夫模型(objectTra-MM)来预测下一个位置。objectTra-MM由两个模型组成：基于用户相似度聚类的马尔科夫模型(object-MM)和基于轨迹相似度聚类的马尔科夫模型(tra-MM)。object-MM分析了对象的空间位置特性,并且将具有相似特性的对象进行聚类。对于每个类,object-MM利用类中对象的轨迹训练变阶马尔科夫模型。tra-MM基于给定的相似度测度对轨迹进行聚类,然后利用每个类中的轨迹训练变阶马尔科夫模型。最后,本文探究了轨迹背后隐含的意图来更好地理解人们的移动模式。为此,我们提出了一种新的概率模型-(Intention of Movement:IoM)来建模轨迹的生成过程。IoM是基于三种重要的现象提出的：(1)轨迹的隐含意图是由轨迹中的位置序列展示的：(2)不同的用户通常有不同的意图；(3)轨迹存在周期性并且在不同的时间段也是不同的。因此,我们通过挖掘位置序列、对象以及时间来发现一条轨迹的隐含意图。为了完整性,我们也提出了只考虑部分属性的简化版模型。最后,我们在两个真实的数据集上进行了大量实验,实验结果验证了IoM的有效性。
[Abstract]:With the wide application of positioning equipment and video capture equipment, spatio-temporal trajectory data acquisition becomes more and more convenient. For example, people on social networking sites (such as Foursquare) and friends to share their location, which has a large amount of data. In another example, the vehicle will be traffic surveillance cameras on the road of the city capture, then using optical character recognition is processed into structured records. Attendance data and traffic data are typical of trajectory data, they all contain objects, location and time of the three kinds of attributes. These data contains a huge value in the city, calculation, path planning, location prediction areas such as the emergence of a large number of the results of this paper. The spatio-temporal trajectory is studied. Firstly, the issues related to mining, prediction of the history of the user's next location problem based on Application in practical. In advance know the user's next location, can recommend more reasonable driving route and recommended to the user will arrive in advertising information position. In order to predict the next position, first proposed the global Markoff model (GMM) and Markoff (PMM.GMM) model using all of the user's trajectory found collective mobile mode; PMM using each user's own personal mobile trajectory found mode. The two models by linear regression combined to produce the final predictor (NLPMM). In addition, because of the time factor on the user's mobile effect model, we seek to join the time factor in the model to improve the prediction accuracy of our proposed. Three methods of clustering in different period trajectory, and training a more refined model to predict the next position. Further, this paper presents a fusion of user similarity Markoff model and trajectory similarity (objectTra-MM) to predict the next position of.ObjectTra-MM consist of two models: the Markoff model of user similarity based clustering (object-MM) and the Markoff model trajectory similarity clustering based on.Object-MM (tra-MM) analysis of the spatial feature of object, and the object will have similar characteristics for each cluster. Class, using the object-MM class object trajectory training variable order Markoff model.Tra-MM similarity measure for clustering based on the given trajectory, and then use each class in the training trajectory of variable order Markoff model. Finally, this paper explores the path behind the intention to better understand the moving mode of people. Therefore, we propose a a new probability model (Intention of Movement:IoM).IoM generation process modeling trajectory is based on three kinds of heavy Put forward to the phenomenon: (1) the trajectory implied intention was demonstrated by the trajectory position sequence: (2) different users usually have different intention; (3) the presence of periodic trajectories and in different time periods is different. Therefore, we dug through position sequence, object and time find a path implied intentions. For completeness, we also propose a simplified version of the model only consider some attributes. Finally, we conducted extensive experiments on two real datasets. The experimental results verify the effectiveness of the IoM.

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP311.13

【相似文献】