CAJ | 학술논문

当前的姿态表示的行为识别方法通常对姿态的准确性做了很强的假设，而当姿态分析不精确时，这些现有方法的识别效果不佳。提出了一种低维的、鲁棒的基于关键肢体角度直方图的人体姿态特征描述子，用于将整个动作视频映射成一个特征向量。同时，还在特征向量中引入共生模型，用以表示肢体间的关联性。最后，设计了分层的SVM分类器，第1层主要用于选择高判别力的肢体作为关键肢体，第2层则利用关键肢体的角度直方图并作为特征向量，进行行为识别。实验结果表明，基于关键肢体角度直方图的动作特征具有较好的判别能力，能更好地区分相似动作，并最终取得了更好的识别效果。
당전적자태표시적행위식별방법통상대자태적준학성주료흔강적가설，이당자태분석불정학시，저사현유방법적식별효과불가。제출료일충저유적、로봉적기우관건지체각도직방도적인체자태특정묘술자，용우장정개동작시빈영사성일개특정향량。동시，환재특정향량중인입공생모형，용이표시지체간적관련성。최후，설계료분층적SVM분류기，제1층주요용우선택고판별력적지체작위관건지체，제2층칙이용관건지체적각도직방도병작위특정향량，진행행위식별。실험결과표명，기우관건지체각도직방도적동작특정구유교호적판별능력，능경호지구분상사동작，병최종취득료경호적식별효과。
The current pose-based methods usually make a strong assumption for the accuracy of pose , but when the pose analysis is not precise , these methods cannot achieve satisfying results of recognition .Therefore , this paper proposed a low-dimensional and robust descriptor on the gesture feature of the human body based on the angle histo -gram of key limbs , which is used to map the entire action video into an feature vector .A co-occurrence model is in-troduced into the feature vector for expressing the relationship among limbs .Finally, a two-layer support vector ma-chine ( SVM) classifier is designed .The first layer is used to select highly discriminative limbs as key limbs and the second layer takes angle histogram of key limbs as the feature vector for action recognition .Experiment results dem-onstrated that the action feature based on angle histogram of key limbs has excellent judgment ability , may properly distinguish similar actions and achieve better recognition effect .