基于运动轨迹聚类的监控视频摘要算法  

Surveillance video summarization algorithm based on motion trajectory clustering

作　　者：李大湘 朱志宇[1] 刘颖 LI Da-xiang;ZHU Zhi-yu;LIU Ying(College of Communication and Information Engineering,Xi’an University of Posts and Telecommunications,Xi’an 710121,China;Key Laboratory of Electronic Information Application Technology forScene Investigation of Ministry of Public Security,Xi’an University of Posts andTelecommunications,Xi’an 710121,China)

机构地区：[1]西安邮电大学通信与信息工程学院,陕西西安710121 [2]西安邮电大学电子信息现场勘验应用技术公安部重点实验室,陕西西安710121

出　　处：《计算机工程与设计》2019年第6期1611-1615,共5页Computer Engineering and Design

基　　金：国家自然科学基金项目(61571361、61102095);陕西省国际合作交流基金项目(2017KW-013);公安部科技强警基金项目(2014GABJC022);中国博士后科学基金项目(2013M542386)

摘　　要：针对监控视频存在大量冗余而不利于检索与传输等问题,提出一种基于运动目标检测和轨迹聚类组合的动态视频摘要算法。采用视觉背景提取子(VIBE)进行运动目标检测,采用运动区重叠判别的方法提取目标运动轨迹;根据视频中所有运动目标的空间位置信息,提出轨迹聚类方法对运动轨迹进行聚类分组;对每个分组的运动轨迹分别建立能量优化模型,采用量子行为粒子群优化算法(QPSO)求解能量方程最小解,将运动轨迹进行重新组合,生成摘要视频。实验结果表明,该算法可将视频浓缩成指定长度的摘要视频,不易发生目标信息丢失或目标重叠,轨迹优化组合性能较传统模拟退火算法提升10%以上。Focusing on the problems of large amount of redundancy in surveillance video and retrieval and transmission failure, a video summarization algorithm based on the combination of moving object detection and trajectory clustering optimization was proposed. The visual background extractor (VIBE) was used to detect the moving target, and the method of overlap judgment of the moving area was used to extract the target motion trajectory. According to the spatial position information of all the moving targets in the video, a trajectory clustering method was proposed to cluster the motion trajectories. An energy optimization model was established for the motion trajectories of each group, and a quantum behavior particle swarm optimization algorithm (QPSO) was used to solve the energy equation minimum solution, and the motion trajectories were recombined to generate a summary video. Experimental results show that the proposed algorithm condenses the video into a summary video of a specified length. It is not easy to lose the target information or overlap the target. The trajectory optimization performance is more than 10% higher than that of the simulated annealing algorithm.

关 键 词：视频摘要 运动目标检测 目标跟踪 运动轨迹聚类 能量最优化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]