基于步态识别的人体目标检测与跟踪

发布时间：2018-05-26 08:26

本文选题：视频监控 + 目标检测　；参考：《北京交通大学》2017年硕士论文

【摘要】：随着计算机技术的发展,智能视频监控技术在民用、军事、航空、医疗和人机交互等众多领域,都有广泛的应用,其重要性也越来越凸显。运动目标的跟踪是计算机视觉领域的一个重要分支,目前,基于视频的目标跟踪技术帮助人们解决了很多问题,但是仍然存在一些不足,提升视频监控的检测效果和跟踪效率,是一个很重要、很值得研究的课题。步态识别作为一种远距离身份判别方式,具有非侵犯性、难以隐藏和难以伪装的特性,鉴于此,将步态识别技术引入到目标跟踪中,实现身份识别,然后进行定位跟踪,在安防、刑事破案等领域具有重要应用和意义。本文针对摄像机固定情况下录制的视频,首先对感兴趣目标区域进行检测,然后通过提取目标的步态特征进行身份识别,最后选定目标实现跟踪。论文工作涉及到了目标的检测、人体步态识别和目标跟踪三个主要部分。论文的主要工作包括以下几部分:(1)在目标检测方面,研究了多种经典的目标检测算法,进行了实验比较分析,综合这几种方法的优缺点,在Vibe算法的基础上进行改进,结合人体体型特征提出了一种PVibe(ProportionVibe)算法,对人体目标进行检测。在简单环境和复杂环境中进行人体目标检测实验对比,实验结果表明,该算法与其他检测算法相比,对于动态的环境具有较好的适应能力,能够很好的提取运动目标轮廓,区分人体目标与非人体目标,具有较好的检测效果。(2)在步态识别方面,使用目标检测分割算法从步态序列中得到步态剪影,提取人体轮廓,将质心到轮廓边缘各像素点之间的距离作为步态特征。利用BP神经网络分类算法对步态样本中各序列的特征数据进行训练识别,从人体步态的不同角度进行实验对比,最高识别率可达到88.33%。该算法与常用的最近邻分类识别算法相比,识别率明显提高,证明了算法的高效性。(3)目标跟踪方面,在对目标进行了步态识别的基础上,选定目标,确定感兴趣区域,并进行跟踪。研究了几种主流的跟踪算法,本文在研究和分析了光流跟踪算法的优点和不足的基础上,结合运动矢量估计提出一种金字塔LK-MVE(LK-Motionvectorestimation)算法,对人体目标进行跟踪。该算法与主流的跟踪算法进行实验对比,在目标颜色相近和出现遮挡的情况下,取得了较好的跟踪效果,而且在跟踪速度上,改进后的算法有明显的提高。
[Abstract]:With the development of computer technology, intelligent video surveillance technology has been widely used in many fields, such as civil, military, aviation, medical and human-computer interaction, and its importance is becoming more and more prominent. Moving target tracking is an important branch in the field of computer vision. At present, video based target tracking technology has helped people solve many problems, but there are still some shortcomings to improve the detection effect and tracking efficiency of video surveillance. Is a very important, very worthy of study of the subject. Gait recognition, as a long distance identification method, is noninvasive, difficult to hide and difficult to camouflage. In view of this, gait recognition technology is introduced to target tracking to realize identification, and then to locate and track. In the security, criminal detection and other areas of important application and significance. In this paper, firstly, the region of interest is detected, then the gait feature of the target is extracted for identification. Finally, the target is selected for tracking. The work of this paper involves three main parts: target detection, human gait recognition and target tracking. The main work of this paper includes the following parts: 1) in the aspect of target detection, we study various classical target detection algorithms, compare and analyze the experiments, synthesize the advantages and disadvantages of these methods, and improve them on the basis of Vibe algorithm. In this paper, a PVibe-ProportionVibe-based algorithm is proposed to detect human targets. The experiments of human body target detection in simple environment and complex environment show that the algorithm has better adaptability to dynamic environment than other detection algorithms, and can extract the contour of moving target well. In gait recognition, gait segmentation algorithm is used to get gait silhouette from gait sequence and extract human contour. The distance between the centroid and the pixels on the edge of the contour is taken as a gait feature. BP neural network classification algorithm is used to train and recognize the characteristic data of each sequence in gait samples. The highest recognition rate can reach 88.33 from different points of view of human gait. Compared with the nearest neighbor classification algorithm, the recognition rate of this algorithm is obviously improved. It is proved that the algorithm is efficient. On the basis of gait recognition, the target is selected and the region of interest is determined. And follow up. Based on the research and analysis of the advantages and disadvantages of optical flow tracking algorithm, a pyramid LK-MVEV LK-Motional moving to timing (LK-MVEV) algorithm is proposed in this paper to track human body targets. Compared with the mainstream tracking algorithm, the proposed algorithm achieves good tracking results in the case of similar color and occlusion, and the improved algorithm has obvious improvement in tracking speed.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】