低分辨率监控视频下的行人再识别方法研究

发布时间：2018-03-16 23:08

本文选题：行人再识别　切入点：行人检测　出处：《中国海洋大学》2015年硕士论文　论文类型：学位论文

【摘要】：行人视频监控已经随处可见,无论是机关单位、大型公共场所,还是私人公寓、小型超市,都布控了各种摄像头。虽然监控设备已经普及,但大多数仍然依赖于人力查看。经常是一个操作人员,或者几个操作人员在很多显示器前逐一查看着是否有异常情况发生。即便是警方查案也是花费大量警力资源逐一查看可疑范围内的视频。智能视频监控还不能达到广泛应用的程度。再加之大规模分布式多摄像机系统迅速膨胀,视觉数据复杂而庞大,简单依靠人力查看显然遥不可及,智能行人识别技术必须紧跟步伐。行人再识别正是面向多摄像头系统监控问题。行人再识别指对于具有时间差距和距离差距、分布在不同位置的摄像头场景下的一个行人或者一组行人,进行视觉上的匹配,以判定是否为同一人,或者给出其出现的所有场景。然而,不同场景下相机角度变化、环境变化、行人视觉外观变化、姿势变化等,造成了行人再识别各个方面、各个研究阶段的困难。本文对行人再识别问题的各个阶段进行了分析研究。首先针对行人检测问题,本文提出了一种基于候选窗口(proposals)的HOG(Histograms of Oriented Gradients)行人检测方法,快速检测行人。然后本文提出了一种改进的基于SLIC的局部-全局超像素图像分割算法,实现行人的准确分割提取,行人提取的准确程度直接影响行人再识别的结果。同时对行人特征空间转换以及匹配算法进行研究,使用CMC评价方法评价行人再识别总体效果。本文实验用到的视频数据来自于实验大楼的监控视频。实验验证了本文提出的基于候选窗口的HOG行人检测算法比现有HOG检测速度快。通过改进的SLIC分割算法,对行人检测的行人图像进行分割,提取行人。最后提取行人特征,运用匹配算法进行实验,得出评价结果。通过本文实验验证,改进了行人再识别总体效果。
[Abstract]:Pedestrian video surveillance has been seen everywhere, whether in government units, large public places, private apartments, small supermarkets, all kinds of cameras. Although surveillance equipment has been widely used, But most still rely on manual viewing. Often an operator, Or several operators in front of a lot of monitors to see if there are any anomalies. Even the police investigation is also a lot of police resources to look at the suspicious range of video. Intelligent video surveillance can not reach. To the extent of widespread application, coupled with the rapid expansion of large-scale distributed multi-camera systems, Visual data is complex and huge, and simply relying on manpower to view it is clearly out of reach. Intelligent pedestrian recognition technology must keep up with the pace. Pedestrian reidentification is a multi-camera system surveillance problem. Pedestrian recognition refers to the time gap and distance gap. A pedestrian or group of pedestrians in different locations of a camera scene performs a visual match to determine whether it is the same person, or to give all the scenes in which it appears. However, the camera angle varies from scene to scene. The changes of environment, visual appearance and posture of pedestrians cause difficulties in every aspect and research stage of pedestrian rerecognition. In this paper, every stage of pedestrian rerecognition is analyzed and studied. First of all, the pedestrian detection problem is studied. In this paper, a new pedestrian detection method based on candidate window proposed by HOG(Histograms of Oriented radientsis proposed to detect pedestrians quickly. Then, an improved local and global super-pixel image segmentation algorithm based on SLIC is proposed to realize accurate pedestrian segmentation and extraction. The accuracy of pedestrian extraction directly affects the result of pedestrian recognition. At the same time, the conversion of pedestrian feature space and the matching algorithm are studied. CMC evaluation method is used to evaluate the overall effect of pedestrian rerecognition. The video data used in this experiment come from the surveillance video of the experimental building. The experimental results show that the proposed HOG pedestrian detection algorithm based on candidate windows is better than the existing one. HOG detection speed. Through the improved SLIC segmentation algorithm, The pedestrian image is segmented to extract the pedestrian. Finally, the pedestrian feature is extracted, and the evaluation result is obtained by using the matching algorithm. The overall effect of pedestrian re-recognition is improved by the experiment in this paper.
【学位授予单位】：中国海洋大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41

【相似文献】