监控视频中基于运动目标显著性的关键帧提取方法  

Keyframe extraction method for surveillance video based on saliency of moving target

作　　者：胡圆圆[1] 干宗良[1] 陈昌红[1] 崔子冠[1] 刘峰[1] 

机构地区：[1]南京邮电大学江苏省图像处理与图像通信重点实验室,江苏南京210003

出　　处：《南京邮电大学学报（自然科学版）》2016年第1期34-41,共8页Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition

基　　金：国家自然科学基金(61071091;61071166;61172118;61471201);江苏省自然科学基金(BK20130867);江苏省高校自然科学研究重大项目(13KJA510004);江苏省"六大人才高峰"资助计划(2014-DZXX-008);南京邮电大学"1311人才计划"资助项目

摘　　要：文中提出了一种监控视频中基于运动目标显著性的关键帧提取方法。该方法首先对监控视频中的行人进行检测和跟踪,分别提取行人的颜色、纹理与形状3种底层特征并借助肤色模型得到能够凸显人脸区域的肤色置信图,然后将3种特征图像和肤色置信图动态加权融合得到多特征融合图像,最后以行人目标的多特征融合图像以及跟踪结果为参考,选取出目标显著性程度最大的融合图像对应的视频帧为关键帧。实验结果表明,文中提出的关键帧提取算法能够快速掌握监控视频中的行人信息,为视频后处理提供运动目标有效的处理样本。A keyframe extraction method in surveillance video based on saliency of moving target is proposed. After pedestrian detecting and tracking,feature maps are constructed based on color,texture and shape of object. The skin confidence map is obtained by skin model,it can help position the facial regions about objects. Then,three feature maps and skin confidence map are dynamic weighted combination into a multi-feature map,including only true attention regions. Finally,with fusion maps and the results about tracking,the frames whose fusion maps can pop out the objects best are selected as keyframes. Results show that the proposed keyframe extraction method can quickly grasp the pedestrian information,making fast retrieval of surveillance video possible. Thus,the algorithm can offer effective samples of objects for video post-processing.

关 键 词：关键帧提取 视频后处理 底层特征 动态加权融合 监控视频 

分 类 号：TN919.8[电子电信—通信与信息系统]