【作者】 王凡；

【导师】 同鸣；

【作者基本信息】 西安电子科技大学 ， 信号与信息处理， 2015， 硕士

【摘要】 人体行为识别与视频分类领域,研究投入逐年增大。近年来,相关产业发展迅猛,如视频监控、智能安防和人机交互等,人体行为识别与视频分类技术在不久的将来还将有着更广阔的市场空间和应用前景。行为识别算法涉及视频数据预处理、特征提取、特征编码、数据降维、聚类分析、模型学习等诸多领域。当下行为识别的研究重心逐渐从底层特征设计转移到模型搭建和中层语义特征提取上。本文从中层语义特征角度,对已有研究成果与现存问题予以比较分析和归纳总结,并做出以下工作:首先,本文概括分析了行为识别中常见的聚类分析算法,针对经典聚类分析算法对初值敏感、聚类中心数目需要人为设定、算法易陷入局部极值、采用欧氏距离不能准确度量特征相似度等问题,给出了一种基于类别的层次聚类分析算法。聚类的目的是为了让特征空间中聚类团簇的类内聚合度高而类间差异度大,为此,算法计算每一类行为在不同聚类中心数目下,类内与类间聚类团簇的相似指标。本文通过分析随聚类中心数目改变,相似度指标的变化趋势,获得类内聚合度、类间差异度的平衡点,这个平衡点对应的聚类中心数目就是本文自适应聚类分析算法得到的聚类中心数目。其次,笔者还构建了判别式聚类分析算法。由于,底层特征、局部特征对视频描述力有限,有效信息被淹没在海量的冗余数据之中,因此对底层信息的总结提炼紧迫且重要。为获取更加有效的中层语义特征,本文提出了判别式聚类分析算法。算法剔除既有聚类团簇中的奇异点,使聚类中心支撑点纯粹度更高。同时,算法对聚类团簇的支撑点数目提出要求,聚类团簇中隶属于某一行为类别的支撑点数目越少,则团簇聚类中心对应该行为的代表性越弱,算法剔除该聚类中心。因此,添加聚类中心支撑点在对应类别的数目约束可以强化聚类中心的表达能力。此外,在迭代过程中,本文算法还将不断削去判别性不足的聚类中心及其支撑点。本文判别式聚类分析算法所求取的聚类中心避免了经典聚类分析算法的弊端,还具有更加优异的判别性、表达性和行为类别纯粹性。再次,本文设计了联合三层语义特征的分类模型,即添加类别约束的隐变量支持向量机(CC-LSVM,Category Constraint Latent Value Support Vector Machine)分类器。为优化底层特征、中层语义特征的识别结果,本文有针对性的提出了高层语义特征。因为,底层特征进行行为识别和视频分类是非语义信息到语义信息的跨越,所以,这种跨越易产生“语义鸿沟”。为连接“语义鸿沟”,论文给出了中层语义特征的构建方法。为建立语义关联,本文提出了能够综合运用三层语义信息的判别模型CC-LSVM,实现多层语义行为分类模型,进一步优化了识别结果。文末,归纳了本论文的重要研究内容,并提出了未来探索的方向。更多还原

【Abstract】 The research, in the field of video classification and human action recognition, is deepen year by year. Related industries, such as video surveillance, intelligent security and human-computer interaction, developed rapidly in the past years. Human action recognition and video classification technology in the near future will have a broader market space. Action recognition algorithm is related to the video data preprocessing,feature extraction, feature coding, data dimensionality reduction, clustering analysis, model learning and many other areas. And today, behavior recognition is focusing on constructing classification model and middle semantic information extraction instead of designing or building the low-level features. In the perspective of middle-level semantic features, this thesis will summarize and analyze the existing research results and existing problems that related, and complete the following work.Firstly, the thesis summarized and analyzed the common clustering analysis algorithm in behavior recognition. In view of some of traditional algorithms have the following problems, such as sensitive to initial setting, the number of cluster centers need artificial setting, the algorithm is easy to fall into local minimum, Euclidean distance metric for some clustering feature similarity is not an accurate measure, this thesis gives a hierarchical clustering analysis algorithm based on the category. The advantages of this clustering algorithm is that, this method allows the feature vectors belonging to the same cluster have a high degree of polymerization in the feature space, but have a high degree differences when feature vectors belonging to different clusters. For that, the algorithm calculates the measurement, which consisting of the polymerization metrics within the class and the dispersion metrics at inter-class, under different cluster number. By analyzing the changing trends of the metrics, consisted of the degree of polymerization within category and the degree of differences among the categories, under different cluster number, we can obtain the balance between the degree of polymerization of features within category and the differences that among the categories. In this way, we can balance the degree of polymerization within category and the degree of differences among the categories and get the number of clustering center adaptive.Secondly, this thesis also built a new framework of discriminative clustering analysisalgorithm. Because 1) the limitation of describing ability of low-level features, local features of the video, 2) the valid information is submerged in a flood of redundant data,hence the summary of the low level information refining is urgent and important. To obtain more effective middle semantic features, we propose a new framework of discriminative clustering analysis algorithm. The proposed clustering framework based on discriminative hierarchical clustering, which will removing the existing cluster clusters singular point, so that it can guarantee the purity of the clustering center anchor. At the same time, the request for the number of features belonging to the clusters anchor is proposed, the fewer the number of support points belonging to a cluster center, the weaker the cluster center corresponding to the behavior, so that the framework will eliminate the poly class center.Thus, the constraints of the number that belonging to the cluster center anchor can strengthen representative of a cluster center. In addition, in an iterative process, this algorithm will continue slashing cluster center anchor which is lack of discrimination.Therefore, discriminative clustering analysis algorithm framework to strike a cluster center not only has richer abundant semantic information, it will also have more excellent discrimination, representative and behavior category purity.Thirdly, we constructed a classification model that can train three-level semantic features at the same time, which add a category constraint CCLSVM(Category Constraint Latent Value Support Vector Machine). To optimize identification results of the low-level features,this thesis proposes high-level semantic features. Because the low-level features to identify and classify the video will occur the semantic information crossing, so it is easy to produce across the "semantic gap". For the establishment of semantic association, connecting the "semantic gap", this thesis presents discriminative model CC-LSVM in this way that three semantic information can be integrated used, multi-level semantic behavior classification model can be realized, and further optimize the recognition results.At the end of the thesis, it summarizes the important research contents of this thesis, and the future direction of exploration in action recognition with mid-level feature.更多还原