最优视频子集与视频时空检索  

Optimum Video Subset and Spatial-Temporal Video Retrieval

作　　者：王美珍 刘学军[1,2,3] 孙开新 王自然 WANG Mei-Zhen;LIU Xue - Jun;SUN Kai-Xin;WANG Zi-Ran(Key Laboratory of Virtual Geographic Environment (Nanjing Normal University), Ministry of Education, Nanjing 210023;State Key Laboratory Cultivation Base of Geographical Environment Evolution (Jiangsu Province), Nanjing 210023;Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023;Nanjing Normal University Taizhou College, Taizhou, Jiangsu 225300)

机构地区：[1]虚拟地理环境教育部重点实验室(南京师范大学),南京210023 [2]江苏省地理环境演化国家重点实验室培育建设点,南京210023 [3]江苏省地理信息资源开发与利用协同创新中心,南京210023 [4]南京师范大学泰州学院,江苏泰州225300

出　　处：《计算机学报》2019年第9期2004-2023,共20页Chinese Journal of Computers

基　　金：国家自然科学基金(41771420,41571389,41401436,41401442);国家重点研发计划(2016YFE0131600);国家高技术研究发展计划(863计划)(2015AA123901);江苏高校优势学科建设工程资助项目;2018年重庆市技术创新与应用示范项目(产业类重大主题专项)(cstc2018jszx-cyztzxX0015)资助;the Sustainable Construction of Advantageous Subjects in Jiangsu Province(Grant No.164320H116)~~

摘　　要：具有时空信息的视频数据爆发式增长给视频数据检索、可视化及其应用带来了严峻的挑战,从海量视频中检索出全面刻画特定目标对象时空信息的最优集合显得尤为重要.针对视频数据具有空间上的聚集性、信息表达的冗余性、视频与检索对象的时序性等特点,该文提出了一种视频时空检索方法,该方法旨在检索最少数量的视频全面刻画检索对象包括时间、空间和方向在内的时空信息.首先,该文分析了视频与目标对象的感知关系;然后,分析了视频感知特征,并提出了感知强度和感知方向的概念,用于被感知目标对象的时空信息特征;在此基础上,给出了最优视频子集的定义,实现了最优视频子集的时空检索方法;最后,通过实验证明了该文方法的可用性.通过两组聚集的视频集合的实验表明,与普通的检索方法相比,(1)本方法能够度量同一感知方向上视频记录目标对象的信息量,并有效选取同一方向上感知度最大的视频,去除冗余视频,因此,能够有效降低结果视频的数据量;(2)本方法保留从各个方向拍摄目标对象的视频,因此,结果视频记录目标对象的信息量与普通检索方法相当;(3)本方法的检索时间与普通方法的检索时间均远小于1 s,本方法检索时间大于普通方法,二者效率差别大小与检索目标对象相关;(4)本方法的效率在视频数据集合确定的情况下,与方向、检索时间、检索对象采样粒度相关,采样方向数量越多,检索时间采样间隔越多,检索对象数量越多,执行时间越长,结果视频数量越多,并且其粒度取值并非越大越好,过多的采样粒度仅仅会增加执行时间,并未获得更多有效的结果视频;(5)本方法方便从视频角度或检索对象进行定量统计与比较;(6)本方法可方便扩展至三维地形与包含障碍物的情况,具有可扩展性.Massive and increasing volume of user-generated videos has brought a major challenge to efficient video query and further applications, such as visualization, time and space analysis. It is important to retrieve a video set that can depict comprehensive spatial-temporal information of the target object. Generally user-generated videos often record the same geographic objects in different directions, and consequently they have redundancy and complementary in space and time. However,current spatial-temporal video retrieval methods only focused on querying the videos according to location, trajectory or/and field of view. In this paper a new spatial-temporal video retrieval method was proposed. It aims to answer “what is a better video set to describe the spatial and temporal information for a geographic object, and how to acquire this video set.” In this paper the better video set is called optimum video set. Firstly perception of the relationship between the video and the target object including the sensing direction and sensing dimension is analyzed. Secondly, the sensing intensity of a geographic object, cell sensing intensity and multi-direction sensing intensity are put forward to quantify the level of videos to record target object’s spatial and temporal information. Videos involved into cell sensing intensity are the candidate ones for the optimum video set. And the optimum video set is defined as videos which have the highest sensing intensity in corresponding cells. The characteristic of the optimum video set is also discussed. Then the method for computing multi-direction sensing intensity and the best subset of videos is proposed. Finally, the experiments verified the proposed method. Experiment results with two aggregated video sets showed that our method has the following advantages compared with the normal spatial-temporal video retrieval method:(1) It can measure the level of videos to record target object’s spatial and temporal information in the same direction, and effectively select the be

关 键 词：视频检索 时空检索 最优视频子集 感知强度 感知方向 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]