CAJ | 학술논문

针对传统行为识别方法仅利用底层特征识别的不足，提出了一种将动作属性与贝叶斯网络相结合的行为识别方法。首先，提取视频中的时空兴趣点及其3 D-SIFT特征描述符，用词袋的方法建立时空词典对视频序列进行表示；然后，利用底层特征训练属性分类器，构造由底层特征到高层特征的映射，将底层特征样本经过属性分类器后得到行为-属性的样本信息，并采用 MAP(最大后验概率)准则学习贝叶斯网络结构，从而建立一种基于属性贝叶斯网络的行为识别模型。实验结果表明该模型能有效地进行行为识别。
침대전통행위식별방법부이용저층특정식별적불족，제출료일충장동작속성여패협사망락상결합적행위식별방법。수선，제취시빈중적시공흥취점급기3 D-SIFT특정묘술부，용사대적방법건립시공사전대시빈서렬진행표시；연후，이용저층특정훈련속성분류기，구조유저층특정도고층특정적영사，장저층특정양본경과속성분류기후득도행위-속성적양본신식，병채용 MAP(최대후험개솔)준칙학습패협사망락결구，종이건립일충기우속성패협사망락적행위식별모형。실험결과표명해모형능유효지진행행위식별。
Due to defect of only using low-level features in the traditional recognition methods,this paper proposes a novel method on action recognition by combining Bayesian Network model with high-level se-mantic concept (human action attribute).Firstly,we have extracted spatio-temporal interest points and 3D-SIFT descriptors around each interest point in the videos.Bag of words as low-level features have been used to describe these videos before building the proj ection from low-level features to high-level features through trained attribute-classifiers.Finally,the Attribute-Bayesian network structure has been studied based on Maximum a Posterior Probability (MAP)mechanism.The experimental results illustrate that the model is effective on action recognition.