【作者】 张晶晶；

【导师】 方贤勇；

【作者基本信息】 安徽大学 ， 计算机应用技术， 2018， 硕士

【摘要】 生物特征技术发展,使得计算机可以通过提取步态特征进行步态识别和步态异常检测。但是监控视频里人群之间的相互遮挡会给前景实例分割和跟踪带来很大困难,现有运动目标提取方法无法针对单个目标分割,而通用人体实例分割方法需要大量的训练数据,且仍然存在分割不完整和误分割等问题。而现有技术之所以无法很好地解决这个问题,是因为方法中没有考虑到每个独立的目标在视频下会具有运动的连续性和时空结构的一致性,所以利用这些特性本文提出了面向步态识别的显著前景分割算法。本文主要包括以下几部分内容:首先,介绍了面向步态识别的显著前景分割算法研究中涉及的相关理论工作。主要从运动目标提取、光流场计算、显著性检测、超像素分割和步态识别这五个方面进行简单介绍,说明了这些理论在我们后续工作的适用性和本文中具体使用的算法。其次,针对前人没有考虑视频中移动目标的运动趋势和显著性差异的问题,本文提出了基于运动持续性和显著性结构约束显著前景分割的步态识别算法。本文将视频序列中运动目标的运动持续性和显著性结构一致性引入,实现基于此两者约束的一种将光流聚类、显著性检测相结合的显著前景分割方法。对于运动连续性,我们提出基于光流区域聚类的前景目标概率计算策略,即基于光流特征对区域进行聚类并以区域面积大小为权重计算前景概率,对于显著性结构一致性,则利用显著性检测结果,获得显著性最强的前景目标区域;最后,利用能量最小化模型对具有完整轮廓的显著目标前景进行像素级别上的能量约束优化,从而实现显著目标前景分割,并在此基础上实现步态识别。最后,针对前面所述方法中的显著型结构约束中,没有具体的考虑人体结构的独有特征,提出了基于粗轮廓多特征组合投票显著前景分割算法,并结合基于块的纹理梯度统计自阴影去除方法,更好的完成步态识别。为了提取更准确的运动区域,减除自阴影的影响,在这里本文还提出了一种基于块的纹理梯度统计自阴影去除方法,可以很大程度上去除运动目标脚底阴影区域。人体结构特征如头、四肢、躯干的相对位置,各部分颜色纹理特征保持相对稳定变化较少等,其中,基于粗轮廓多特征组合投票是针对人体轮廓内时空一致性的约束特征,以超像素分割得到的超像素块为基本单位,在时间上,以超像素块的纹理特征用学习到的模板对待分割区域进行投票;在空间上,以超像素块的相邻区域特征用学习到的模板对待分割区域进行投票,这样来获得与模板最相近的目标区域。将这两个方法与前面的运动连续性和显著性结构约束相结合,再进行步态识别,实验结果表明,基于粗轮廓多特征组合投票显著前景分割的步态识别得到了更好的识别效果。更多还原

【Abstract】 The advance of biometrics technology enables the computer to detect gait and gait abnormality on intelligent monitoring.However,the interception of crowds in the surveillance video will bring great difficulties to the foreground segmentation and tracking.The existing moving object extraction methods still have difficulty in robust segmenting the object,without considering that each independent target will have the motion continuity and the space-time structural consistency under the video.Aiming for this shortcoming,this thesis does the following work:First,the existing d work on foreground segmentation algorithm for gait recognition is introduced briefly.Five topics are presented:moving object extraction,optical flow field calculation,saliency detection,super pixel segmentation and gait recognition.Secondly,a novel gait recognition algorithm based on the constraints of both motion continuity and salient structure of foreground.The method uses optical flow clustering and saliency detection for both motion continuity and salient structure of foreground,which not considered in existing work.For the motion continuity,an estimation strategy for the foreground probability is introduced based on the clustering of optical flows.It clusters the regions based on optical flow and computes the foreground probability based on the area size.For the salient structural consistency,a scheme based on the saliency detection is presented.A energy minimization model is introduced by integrating the two strategies for a salient foreground area and consequently a robust gait recognition can be resulted.Finally,a novel gain recognition method based on salient foreground segmentation using a coarse-contour multi-feature voting scheme is proposed.First,a block-based gradient self-shadowing algorithm for texture statistics is proposed here,which can largely remove the shadow area of the sole of moving target.This algorithm extracts a more accurate motion area and eliminates soft shadows.Second,a salient foreground segmentation is obtained with a rough-contour based multi-feature voting which includes the spatial and temporal coherence constraints according to the human body features.Integrating these two methods into the former gait recognition method using constraints of motion continuity and saliency,a new gait recognition method is obtained.Experiments show that this novel method obtains a better recognition effect than the previous methods.更多还原