人的行为分析是计算机视觉领域的一个重要研究方向，它在智能监控系统、高级用户接口、人的运动分析、虚拟现实等领域有着广泛的应用。行为分析的最终目标是要通过对行为特征数据的分析以获取行为的语义描述与理解，而行为建模和识别是达到上述目标的一个关键步骤。本文从半/无监督学习方法的角度，对行为建模和识别展开了研究，主要工作有： 1．提出了一种基于时空图像的行为建模与识别方法。该方法首先将视频序列转化为时空图像，并从时空图像中提取Gabor纹理特征，然后在已知类别数的情况下，利用基于动态时间归整的谱聚类方法自动地标记样本，最后建立行为的模板。这种方法简单易行，是一种无参建模方法，不需要大量的训练样本，不需要跟踪人体的各部位，而且对前景提取的精度没有苛刻的要求。 2．提出了一种基于半监督学习的行为建模与异常检测方法。该方法结合基于动态时间归整的谱聚类和最大后验自适应两种技术，能够在大样本的情况下自动地选择正常行为模式的种类和样本来建立正常行为的隐马尔可夫模型，能够在少样本的情况下避免欠学习问题，建立可靠的异常行为的隐马尔可夫模型。 3．提出了一种行为模型的在线学习方法：MAPACo-Training。传统的Co-Training算法是一种离线学习方法，对“每个特征空间分量都能足够地学习目标概念”这一假设条件有着苛刻的要求。本文方法结合了Co-Training与最大后验自适应两种技术，不仅能够克服传统方法对这个假设条件的苛刻要求这一问题，而且还能够用于多类在线学习。; Behavior analysis is driven by a wide range of applications, such as visual surveil- ance, advanced user interface, video meeting, behavior based video index and retrieval, medical diagnosis and so on, and it is an active research topic in computer vision.The aim of behavior analysis is to get semantic descriptions and understandin- gs. To this end, a key step is to model and recognize behaviors. Based on semi-super- vised/unsupervised learning methodology, this thesis is focused on behavior modeling and recognition. The main contributions are: 1. A novel method based on space-time image features is proposed for automatic behavior modeling and recognition. The method is composed of the following three steps: (1) Video sequences are converted into a space-time image, and then the features are extracted by Gabor filtering; (2) With the known number of behavior types, an unsupervised technique based on dynamic time warping is used to determine the groups of different behaviors; (3) The behavior templates are built according to these groups, and are used for the recognition of behaviors. The method does not need to track the body parts and can model the behaviors without any prior knowledge. It is a non-parametric method and only needs a small number of samples to build template. 2. A simple and efficient method based on semi-supervised learning technique is proposed for behavior modeling and abnormality detection. Combining dynamic time warping based spectral clustering and maximum a posteriori adaptation, this method can automatically select the number of normal behavior pattern types and samples from the training dataset to build the hidden markov models of normal behaviors, and can effectively avoid the running risk of over-fitting when the hidden markov models of abnormal behaviors are learned from sparse data. 3. The traditional co-training algorithm, which needs a great number of unlabeled examples in advance and then trains classifiers by iterative learning approach, is not suitable for online learning of classifiers. In addition, it requires that each of feature view component in the feature space is sufficient for learning the target concepts. To overcome these barriers, we propose a novel semi-supervised learning algorithm, called MAPACo-Training, by combining the co-training with the principle of maximum a posteriori adaptation. This MAPACo-Training algorithm is an online multi-class learning algorithm and overcomes the requirement. It has been successfully applied to online learning of behaviors modeled by hidden markov model.
修改评论