摘要
针对动作识别算法中构建局部特征间关系不准确、先验知识少的问题,采用一种隐马尔科夫模板模型(HMT),用于从视频中识别人的动作。通过隐马尔科夫模型的状态转移概率,描述动作的各个姿态间的关系。通过模板匹配的方式来计算观测值到隐含状态之间的似然函数(因此称为隐马尔科夫模板模型),模板是活动的,可以克服目标几何外观的变化。隐马尔科夫模版模型的参数包括所有关键姿态模板的参数和这些关键姿态之间的转移概率。这些参数可以通过"期望最大化"算法从训练数据中学习。实验结果表明,所提出的方法在识别准确率上有明显优势。
In order to resolve problems such as incorrect constructing relationship of local feathers and less prior knowledge,a hidden Markov template( HMT) model was presented for recognizing human activities from video clips.Transition probabilities of hidden Markov model describe the relationship of key poses.Each key pose in the HMT model has a video template( hence the name) that is used to measure the likelihood between an observation and the key pose by template matching.The templates are active,so they are robust to changes in appearance.The parameters of a HMT model include the video templates of all key poses as well as the transition probabilities between them.These parameters can be learned with a modified version of the Expectation-Maximization( EM) algorithm.Experimental results demonstrate that the HMT model achieves better on recognition precision.
出处
《武汉理工大学学报（信息与管理工程版）》
CAS
2013年第6期789-793,共5页
Journal of Wuhan University of Technology：Information & Management Engineering
基金
国家"863"计划基金资助项目(2012AA041203)
