【作者】 马华兵；

【导师】 张二虎；

【作者基本信息】 西安理工大学 ， 信号与信息处理， 2008， 硕士

【摘要】 在对安全监控要求越来越高的今天,生物特征识别技术以其良好的安全性,有效性和不易取代性,受到人们的重视。而步态识别是利用人的走路姿势来进行生物识别的一种方法,可以在被监控对象毫无察觉的情况下对其进行检测。以其非接触、远距离识别的特点备受人们的青睐。本文研究了基于机器学习和多视角信息融合的步态识别技术,主要研究了人体步态特征描述,利用机器学习进行降维,单视角下步态识别和多视角下信息融合的步态识别以及搭建在线步态识别系统等相关问题。针对采用背景减除法进行背景提取时0度视角下人体部分区域被分割成背景的问题,在步态能量图像的基础上,提出了动态步态能量图像,较好的解决了这一问题。并且在90度和45度视角下,由于动态步态能量图像关注的是人体区域相对变化的部分,避免了相对不变化的部分带来的负面影响,从而也取得了较好的识别结果。对于背包情况,提出一种动态腿部能量图像,抛弃上半身的特征信息,只利用腿部动态变化信息进行识别,也取得了较好的结果。在特征描述阶段,采用机器学习的方法进行降维,利用流行学习中的局部保留投影法(LPP),采用PCA+LPP的方法利用很少的维数保留了很好的特征。针对不同视角下人体的步态序列,利用信息融合的方法,用投票法和D-S证据理论法对不同视角的识别信息进行融合,有效地解决了单个视角对特征描述不完全的问题,在中科院步态库NPLSR和CASIA库上进行了测试,取得了较理想的结果。在上述理论的基础上,尝试建立了在线步态识别系统,采用动态能量图像作为特征,利用局部保留投影进行降维,利用多视角融合的方法进行识别,在实际测试环境下,取得了较好的结果。更多还原

【Abstract】 Today, the security is becomes more important, researchers take more attention to biometric recognition with the favorable security, validity and hard replace. Gait recognition attract the eyeballs of more and more researchers because of non-contact and far distance, which make use of the human walking pattern to recognition and identify, and executed backdoor.In this paper, gait recognition technique based on machine learning and fusion of multi-view is researched. And gait feature description, dimensionality reduction with manifold learning and single view and fusion of multiple views gait recognition are the mainly content. Firstly we give a new gait feature describe method—dynamic gait energy image based on the method gait energy image, due to some parts of human body are transacted background at direction 0, and which avoid the negative effect of the non-moving parts at direction 45 and 90. When people get the knapsack, we only use the information of the leg, but throw away the information of the upside. Secondly we use the method principal component analysis (PCA) + locality preserving projections (LPP), get the better recognition result with the lower dimension. Thirdly the fusion method is used with the gait image sequence of three views (direction 0, direction 45 and direction 90). It solves the problem that feature is not enough in single view. Lastly the method above is tested in three gait databases (NPLSR, CASIA and timer shoot), and the results indicate that gait recognition using multiple views information fusion can obtain higher recognition rate than any other single view.Meanwhile, a online gait recognition system is built, the method above is used, this is a attempt that gait recognition used in practical process, and it is a tested flat for gait recognition.更多还原