【作者】 钱惠敏；

【导师】 王执铨； 茅耀斌；

【作者基本信息】 南京理工大学 ， 控制科学与工程， 2009， 博士

【摘要】 基于视觉的人体运动分析是指对视频中的运动人体进行检测、识别和跟踪,并理解和描述人体行为,它是计算机视觉领域中一个新兴的研究方向。人体运动分析系统通常涉及图像预处理、运动目标检测和识别、运动人体跟踪、人体行为理解与描述等几个主要研究内容。目前在运动目标检测、识别及跟踪上得到相对成熟的研究成果,而在人体行为理解与描述方面,仍然存在大量问题亟待解决。本文就人体运动分析系统中的图像平滑、跟踪算法的尺度自适应性,以及人体行为理解等问题展开研究,取得的主要成果包括：(1)提出了一种基于视觉掩蔽效应的各向异性扩散系数方程,改进了各向异性扩散图像平滑算法,能有效滤除人体运动序列图像中的噪声。比较了现有几种基于各向异性扩散方程的平滑算法的性能,实验结果表明,本文方法不仅使恢复图像质量更好,信噪比更高,且算法的收敛速度快。(2)由于Mean-Shift算法和粒子滤波算法在目标尺寸变化大时不能准确跟踪,定义了多尺度图像信息量(MSIIM),基于此信息量,提出了一种自动更新跟踪窗口尺寸的目标跟踪算法。实验结果表明,本文改进的跟踪算法,对逐渐增大的运动目标和逐渐减小的运动目标都能实现准确地实时跟踪。(3)提出了一种无监督的人体运动序列分割算法,并对分割出的行为段采用隐马尔可夫模型(HMM)识别。然而,当前多数行为识别算法只关注于识别包含单个行为模式的行为段,忽略了运动序列的分割。本文提出通过检测特征向量的本征维数的突变点实现运动序列的分割,本征维数采用主成分分析(PCA)方法估计。提出了一种基于紧轮廓编码的点集特征描述运动人体。在当前最常用的两个公共数据库上的实验结果表明本文算法的有效性。(4)采用安装在室内墙体上的Panasonic (WV-CW960)摄像头自行拍摄了人体各类行为视频,建立了行为数据库OwnSet,作为后续行为识别和异常检测算法的研究对象。实验场景符合实际监控场景,数据库共包含六种行为：行走、慢跑、坐下、蹲下、跌倒和不动行为(站立、蹲着、坐着)。(5)提出了一个多类行为分类识别系统,包含四个模块：视频采集、运动目标检测与定位、特征提取以及行为分类识别。在运动目标检测与定位模块,采用背景减除法；在特征提取模块,分别提出了运动能量序列(MES)和运动能量图像轮廓编码(CCMEI)特征；在行为识别模块,引入了分层分类的思想,分别提出了基于先验知识的决策树支持向量机多分类器和基于聚类的决策树支持向量机多分类器,且后者比前者更具有通用性。在数据库OwnSet和公共数据库上的实验结果表明本文提出的系统对不完整的运动人体块,严重的阴影,以及各种不同的衣着等都表现出了稳定的识别性能,识别率明显高于已有算法。(6)研究了人体运动分析在智能监控中的应用,提出了家居看护中的日常行为异常检测算法,实现在无人照看的老人、小孩或病人遇到跌倒危险时监控系统给出及时告警。提出了基于组合分类器的跌倒检测算法,从OwnSet数据库的六种行为中检测出跌倒；提出了基于支持向量机分类器的跌倒检测算法,不仅能检测跌倒行为,还能区分是行进中发生的跌倒行为还是原地跌倒行为,从而判断跌倒的危害程度。实验结果表明算法的识别性能较高,且由于算法简单,系统实时性高,易于在实际应用中推广。最后,分析了目前人体运动分析系统中存在的一些困难和亟待解决的问题,指出了进一步的研究方向。更多还原

【Abstract】 Vision-based human motion analysis aims at detecting, identifying and tracking moving people from video sequences, and furthermore understanding and describing human activities. It is one of the most active research topics in computer vision. Human motion analysis sytem usually involves image preprocessing, moving object detection and identification, object tracking, and human behavior understanding and description. Although many research results have been achieved on object detection and tracking, there are still a lot of issues to be settled in behavior understanding and description.In this dissertation, image smoothing, adaptability to scale changes of tracking algorithm and human behavior understanding are considered. The main research achievements include:An anisotropic diffusion coefficient equation based on visual masking effect is proposed, which improves the performance of anisotropic diffusion image smoothing method. The noise in image sequences is removed effectively by the improved smoothing method. It is compared with several existed image smoothing algorithms. Experimental results show that not only the smoothed image by our method is better according to the signal-to-noise ratio and the mean structure similarity, but the convergence speed of our method is faster.The information measurement of multi-scale images in scale space is investigated, based on which an object tracking algorithm with self-updating tracking window is proposed. The algorithm improves the Mean-Shift and partical filtering tracking methods and makes the moving objects with increasing scales or decreasing scales are both tracked exactly. And the improved tracking algorithms also run in real time.An unsupervised temporal segmentation algorithm of human motion sequences is proposed, and then the obtained activity segments are recognized by HMM. However, most of the existed activity recognition algorithms are conducted on the segmented activities. In this dissertation, moving human blobs are described by a new feature which is a compact contour point set. Then temporal segmentation of human motion sequences is achieved by detecting the change point of the intrinsic dimensionality of feature vector, which is estimated by PCA. Experiments on two commonly used public databases demonstrate the effectiveness of our algorithm.Videos consist of multiple types of human actions are collected by a Panasonic camera (WV-CW960) mounted on the wall, and the experimental environment coincides with the real surveillance system. An activity data set called OwnSet is constructed by the collected videos, on which the subsequent activity recognition and abnormality detection algorithms are conducted. There are six kinds of actions in OwnSet, which are walk, jogging, sitting down, squatting down, falling down, and immovability (including stand, squat, and sitting).A system to recognize multiple kinds of activities is presented. It consists of four modules:video collection, moving objects detection and location, feature extraction, and activity classification. Background subtraction is employed to detecting moving persons. Two new features named Motion Energy Sequences (MES) and Contour Code of Motion Energy Image (CCMEI) are extracted. In the activity recognition module, the hierarchical classification thought is introduced. Two support vector machine decision-tree classifiers based on apriori knowledge and clustering results respectively are used to recognize multiple kinds of actions, and the latter classifier is more universal than the former. Experimental results on the OwnSet and a public database show that the recognition accuracy of our algorithm is bigger than other algorithms and it is roubst to incomplete human blob, heavy shadows, as well as human appearance variation.The application of human motion analysis in intelligent surveillance is studied in this paper. Abnormality detection from multiple daily activities is conducted. Alarms will be issued in time when the olders, the children or the patients encounter a fall. An abnormality detection algorithm based on a combined classifier is presented to detect fall from the six types of actions in OwnSet. Since the causes and dangers of on-marching falling down and in-place falling down are usually different, another detection algorithm based on support vector machine is proposed. Experimental results demonstrate that both algorithms have good recognition performance and real-time performance. Therefore, the proposed algorithms are easily extended in real applications.Finally, some difficult problems to be settled in human motion analysis system as well as the future work are listed out.更多还原