【作者】 杨伟；

【导师】 艾廷华；

【作者基本信息】 武汉大学 ， 地图制图学与地理信息工程， 2019， 博士

【摘要】 在大数据智能时代,随着无线传感定位设备的普及使用,移动对象时空轨迹数据呈现爆炸式增长。轨迹大数据包括交通车辆轨迹、人类活动轨迹、动物迁徙轨迹、自然现象轨迹等,具有数量大、类型多、增长快、价值高、众源泛在、时空多维、精细分辨率等优点。由于其巨大的研究应用价值,已有诸多重量级研究成果相继涌现,并广泛服务于智能交通、城市规划、服务推荐、行为分析、环境监测、公共安全、城市计算、社会感知等领域。然而,大数据的“5V”特性给轨迹数据挖掘处理、知识发现带来了新的挑战,使其面临“数据丰富而知识贫乏”或“知识即常识”的困境。为此,建立快速、高效的轨迹数据结构化处理模型方法,对轨迹大数据去粗取精、去伪存真以提取高价值、结构化的时空知识成为迫切解决的问题。论文以时空轨迹数据为研究对象,以结构化道路信息化提取与更新和活动行为建模与场所语义感知为研究问题,开展了基于时空轨迹数据的结构化处理理论模型与技术方法研究。论文研究工作主要包括以下几个方面:1.针对结构化道路信息快速提取问题,提出了一种基于低频车辆轨迹数据的车行道路面、线一体化提取技术方法。首先,提出自适应优选加密方法解决低频车辆轨迹中存在的假轨迹段、高噪音、采样稀疏、轨迹点数量少等问题。然后,引入约束Delaunay三角网及Voronoi图模型建模车辆轨迹数据,建立基于轨迹密度分布差异的边界识别指标,并集成识别指标、轨迹运动特征构建道路边界探测模型。最后,通过探测模型在轨迹几何细节层次上的分类处理,提出基于Delaunay三角网的道路面、线信息提取模型与算法。该方法顾及车辆轨迹在路网空间中的上下文差异特征,适于处理复杂路网结构、多时间跨度、多密度差异情形下的低频车辆轨迹数据,实现了基于“车辆轨迹->道路边界->路网中线->车行地图”数据处理流程的车辆轨迹结构化建模。2.针对路网变化信息快速发现与更新问题,提出了一种基于众源车行轨迹数据的路网变化信息检测与更新技术方法。首先,分析道路变化信息与车辆轨迹间的关联关系,为路网现势性检测与更新提供理论支撑。然后,集成轨迹运动特征、车行语义、车行轨迹与路网关联知识构建道路变化信息检测与识别模型。最后,提出以道路弧度缓冲区为基本更新单元,通过在轨迹几何特征上的精细化处理,在局部范围内探测路网变化、识别道路变化类型,根据变化类型提取变化信息、快速完成路网更新。相比已有变化发现技术,该方法能在更精细粒度上识别道路局部变化、路网方向变化、判别真实变化与语义变化,实现了基于“局部分析->类型推断->变化提取->增量更新”数据处理流程的路网变化检测更新。3.针对中小区域精细道路地图构建问题,提出了一种基于多源多类活动轨迹数据的中小区域路网精细建模与地图重建技术方法。首先,深入分析人类活动轨迹数据特征和中小区域路网精细建模过程,为精细道路数据获取与地图构建提供理论支持。其次,根据“轨迹数据->轨迹转弯点->聚类中心->道路节点”的逻辑关系,提出轨迹转向点探测算法提取转向轨迹点集、ASCDT算法聚类轨迹转向点建模道路节点。最后,根据“轨迹分割->增量融合->路网构建->优化处理”的处理顺序,提出轨迹线分割、轨迹线增量融合、优化处理等系列方法分层构建多专题精细道路地图。该方法能快速获取中小区域路网信息、完成路网几何、拓扑、语义信息的精细建模、构建多主题路网地图,实现了基于“活动轨迹->道路节点->路网拓扑->道路地图”数据处理流程的精细道路地图重建。4.针对活动行为识别与场所语义感知问题,提出了一种基于时空轨迹数据的活动行为建模与语义特征信息提取技术方法。该方法基于轨迹几何、语义特征,在个体层面运用轨迹运动参数建模活动行为、提取典型化行为语义特征;分别运用三角网模型、语义增强等方法在群体聚合层面分析活动行为事件与场所关联耦合特征、提取活动场所位置语义信息。以加油停留行为和慢跑绕圈行为及其场所语义感知为例,初步建立了从“行为模式挖掘”到“场所语义感知”的轨迹语义感知框架。该框架通过结构化分而治之策略,提取结构化几何、语义特征建模活动行为、挖掘行为语义、完成深层次的场所语义理解,实现了基于“时空轨迹->行为建模->活动探测->场所感知”数据处理流程的活动场所挖掘分析。5.通过研发时空轨迹数据挖掘处理原型系统,实现了时空轨迹数据的一系列挖掘处理分析,验证了论文相关模型、算法的有效性和可行性。综上,论文开展了轨迹密度特征上的时空异质性分析、几何计算与空间统计融合的道路信息获取更新、时空上下文的典型化语义特征提取、活动行为事件的场所位置耦合挖掘等相关工作,提出了系列轨迹数据结构化处理模型、算法,实现了轨迹数据从几何细节层次到语义概化层次的建模处理与语义理解。但轨迹大数据建模处理作为一项既复杂又系统的工作,仍存在诸多问题需要深入研究。更多还原

【Abstract】 In the age of big data and artificial intelligence,with the widespread use of wireless sensor positioning equipment,moving objects trajectory data shows an explosive growth.Trajectory data includes vehicle trajectory,human activity trajectory,animal migration trajectory,natural phenomenon trajectory and so on,which has many advantages,such as big data,high value,and crowdsourcing and so on.Due to its huge research and application value,many research achievements have sprung up,and it is widely used in intelligent transportation,urban planning,service recommendation,behavior analysis,environmental monitoring,public safety,urban computing,social sensing and other fields.However,the 5V characteristics of big data bring new challenges to trajectory data mining and knowledge discovery,which makes it face the predicament of “rich data and poor knowledge” or “knowledge is common sense”.Given this problem,establishing a fast and efficient structured processing model and method to extract high-value,structured knowledge from trajectory data has become an urgent problem to be solved.Taking trajectory data as research object and taking road information extraction and update and activity behavior modeling and place semantic perception as research issues,this paper studies the structured processing model and method of trajectory big data.The main research work are listed as follows:1.Aiming at the problem of fast extraction of structured road information,a new method of road boundary and road line extraction based on low frequency vehicle trajectory data is proposed.Firstly,an adaptive interpolation method is presented to solve the problems of high noise and sparse sampling.Then,the constrained Delaunay triangulation and Voronoi model are introduced to model vehicle trajectory data at the level of geometric details,and the road boundary detection model is constructed by integrating the trajectory movement features and boundary identification indicators.Finally,the model and algorithm of road information extraction based on Delaunay triangulation are presented.Experimental results showed that the proposed method is suitable for GPS traces with complex road structure,disparity density and different time interval,and the structured trajectory data modeling based on the data processing flow of “trajectory-road polygon-road line-road map” is realized.2.Aiming at the problem of fast updating of road data,a method of road change information detection and updating based on crowdsourcing vehicle trajectory data is proposed.Firstly,the relationship between road change and trajectory data is deeply analyzed to provide theoretical support for this method.Then,a model of road change information detection and classification is constructed by integrating movement features and semantic information.Finally,after fine processing of the track geometric features,a new method that taking road segment buffer as basic update unit is presented to detect road change information,identify the type of road change and update the road data.Compared with the existing method,our method can identify local changes,direction changes and semantic changes of road network in finer granularity,and road change detection and update based on the data processing flow of “local analysis-type inference-change extraction-incremental update” is realized.3.Aiming at the problem of fine road map construction,a method of fine road information acquisition and map reconstruction based on crowdsourcing trajectory data is proposed.Firstly,the features of activity trajectory data and the process of fine road map reconstruction are analyzed deeply to provide theoretical support for this method.Secondly,according to the relationship of “trajectory data-turning point-cluster center-road node”,the turning point detection algorithm is proposed to extract turning point set,and the ASCDT algorithm is used to cluster turning points to model road node.Finally,according to the processing flow of “trajectory segmentation-incremental fusion-network construction-optimization processing”,a series of methods,such as trajectory segmentation,incremental fusion,optimization processing,are proposed to build fine road map.The presented method can quickly acquire finer road information,complete the fine modeling of road network geometry,topology,and semantic and construct multi-thematic map,and the fine road map reconstruction based on the data processing flow of "trajectory-road node-road edge-road map" is realized.4.Aiming at the problem of activity behavior recognition and place semantic perception,a method of activity behavior modeling and semantic feature information extraction based on spatiotemporal trajectory data is proposed.Based on the geometric and semantic features of track,the method uses trajectory movement parameters to model activity behavior and extract the semantic features of typical behavior at the individual level.Then,Delaunay triangulation model and semantic enhancement method are used to analyze the coupling features between behavior events and places and to extract the location semantic information of places at the aggregation level.Taking refueling stop behavior and jogging circling behavior as cases,the framework of trajectory semantic perception from “behavior pattern mining” to “place semantic perception” is established.The framework extracts structured geometry and semantic features to model activity behavior,identify behavior patterns,mine behavior semantics and complete deep understanding of place semantics through the strategy of structured divide and conquer,and the mining and analysis of activity place based on the data processing process of “spatiotemporal trajectory data-behavior modeling-activity detection-place perception” is realized.5.Lastly,this paper develops a trajectory data mining prototype system.And a series of mining,processing and analysis of spatio-temporal trajectory data are realized,and the validity and feasibility of the related models and algorithms are verified.In summary,this paper carries out some works on spatio-temporal trajectory data mining,including spatio-temporal heterogeneity analysis of track density features,road information acquisition and updating through fusion geometric computing and spatial statistics,typical semantic feature extraction from spatial-temporal context,minging of the coupling features between behavior events and places,etc.And a series of structured processing models and algorithms for trajectory data are presented,and modeling processing and semantic understanding of trajectory data from geometric detail level to semantic generalization level are realized.However,trajectory data processing as a complex and systematic work,there are still many problems to be studied in depth.更多还原