【作者】 胡荣；

【导师】 王宏远；

【作者基本信息】 华中科技大学 ， 通信与信息工程， 2010， 博士

【摘要】 生物特征识别技术是一种利用人的生理或行为特征进行身份识别的技术。随着机场、车站、银行等安全敏感场合对大范围视觉监控系统的需求提升,远距离的身份识别研究近来受到了计算机视觉研究者们的大量关注。脸像、指纹和虹膜等生物特征通常需要近距离或者接触性的感知,因此在实际应用中受到了诸多限制。步态是远距离情况下唯一可以感知的生物特征,因此从视觉监控的观点来看,步态识别具有广泛的应用前景。步态识别是一个新兴的研究领域,它利用人走路期间的行为特征来进行身份识别。围绕这个主题,本文开展了如下几个方面的研究工作：1.步态轮廓提取是步态识别系统的重要步骤,它包括了背景建模和运动分割两个部分。本文采用核高斯模型(Kernel Gaussian Model)来对步态场景进行建模,同时提出了一种高效的高斯核带宽估计方法。步态图像通过背景模型被映射为背景概率图像之后,EM算法联合运动段分析的方法将自动地分割出概率图像中的人体运动区域。实验表明,该方法在复杂场景中的步态轮廓提取效果要明显优于传统的步态轮廓提取方法,并且它的运算效率能够满足动态场景的实时建模要求。2.在步态轮廓较为清晰的情况下,本文采用模型化的方法,通过骨架模型来提取人体的步态识别信息。一种改进的基于距离变换的骨架算法首先被用来提取人体骨架；然后,通过骨架的腐蚀与还原技术从骨架模型中提取人体的躯干和腿部骨架枝；最后,躯干和腿部骨架的周期性运动特征作为步态的识别特征被用来进行个体身份识别。实验结果表明,该方法在室内环境下有着理想的识别成功率。特别在多视角的情况下,该方法可以通过对骨架模型进行视角修正来极大地提高其正确识别率。3.在步态轮廓不甚理想的情况下,本文采用全局的方法,通过对步态的空时数据进行连续的特征子空间学习来提取识别信息。第一次特征子空间学习对步态的频域数据进行主成分分析,并将步态数据转化为周期特征矢量形式；第二次特征子空间学习对步态的周期特征矢量形式继续进行主成分分析加线性判别分析的联合处理,最终将步态数据转化为步态特征矢量。步态特征矢量既包含了人体的形态特征又包含了人体的运动特征,因此具有很强的识别能力。实验结果表明,该方法不仅获得了令人满意的正确识别率,而且还拥有相对较低的计算与存储代价。更多还原

【Abstract】 Biometrics makes use of the physiological or behavioral charactristics of people to authenticate their identities. With the growing need for a full range of visual surveillance and monitoring system in security-sensitive envirenments such as airports, bus stops and banks, human identification at a distance has recently gained increasing interest from computer vision researchers. To operate successfully, the established biometrics such as face, fingerprints and iris usually require proximal sensing or physical contact. However they are hardly applicable at a distance. Fortunately, gait is still visible and can be easily perceived unobtrusively. So, gait is a very attractive modality from the visual surveillance perspective.Gait recognition is a relatively new research area. It aims to seek distinguishable variations between the same actions of walking from different people for the purpose of automatic identity verification. Focusing on this topic, this dissertation mainly includes the following issues:1. Silhouette extraction is an important procedure in gait recognition, which includes two steps:background modeling and motion extraction. In this dissertation, Kernel Gaussian Model is used to model the gait background, and an effective kernel width estimation method is proposed. After gait images are mapped into probability images by the statistical background model, foreground pixels are extracted from these probability images by the EM algrithem and motion slice analysis. Experiment result shows that the proposed method achieves obvious improvement than traditional methods for the complicated scenes. And at the meantime, it can satisfy the requirement for real-time background modeling.2. Gait features are extracted by the model-based method of skeletonization under the situation that silhouette quality is relatively good. First, an improved distance transform method is proposed to extract the raw skeleton; then, the main branches of human body are extracted by the skeleton erosion and restoration technique; lastly, the periodic motion charactaristcs of human main skeleton are treated as the final gait feature to identify people. Experiment result shows that the skeleton feature of gait can achieve highly satisfying performance in indoor environment. Especially for multipul view settings, the recognition rate can be significantly increased by fixing the view angle of skeletons.3. A novel gait recognition method based on spatiotemporal feature extraction is proposed under the situation of bad silhouette quality. In the first subspace learning, the periodic dynamic feature of gait is extracted by Principal Component Analysis and sequence data is represented in the Periodicity Feature Vector form; in the second subspace learning, Principal Component Analysis integrated with Linear Discriminant Analysis are applied to the Periodicity Feature Vector representation of gait and sequence data is compressed into Gait Feature Vector. Gait Feature Vector contains both the shape and the dynamic information of human gait, which shows strong discriminative ability. Experimental result shows that the proposed method achieves highly competitive performance and it takes less storeage and computational cost.更多还原