CAJ | 학술논문

提出了一种利用AdaBoost算法对原始特征进行提取的人体动作识别方法，即从表征大量的训练数据的原始特征中提取出少量的、可分性高的特征样本，进而在保证识别率的基础上，减少了识别的计算量，达到快速、准确的人体动作识别。首先，利用基于时空兴趣点的三维尺度不变特征变换(3D SIFT)描述子对运动视频序列进行表示得到原始特征数据，然后将表征每个视频序列的原始特征通过AdaBoost算法进行提取，筛选出对于识别贡献最大特征数据作为训练样本。最后利用最近邻分类器在现有的KTH数据库上对本文方法进行测试，实验表明在得到较好的识别率的同时，大大提高了识别的速度，验证了所提方法的有效性。
제출료일충이용AdaBoost산법대원시특정진행제취적인체동작식별방법，즉종표정대량적훈련수거적원시특정중제취출소량적、가분성고적특정양본，진이재보증식별솔적기출상，감소료식별적계산량，체도쾌속、준학적인체동작식별。수선，이용기우시공흥취점적삼유척도불변특정변환(3D SIFT)묘술자대운동시빈서렬진행표시득도원시특정수거，연후장표정매개시빈서렬적원시특정통과AdaBoost산법진행제취，사선출대우식별공헌최대특정수거작위훈련양본。최후이용최근린분류기재현유적KTH수거고상대본문방법진행측시，실험표명재득도교호적식별솔적동시，대대제고료식별적속도，험증료소제방법적유효성。
This paper proposed a novel action recognition method based on AdaBoost algorithm. Selecting the most discriminative sample subset from a large amount of raw features of training data,the method re-duces the computational complexity of recognition with high accuracy,and realizes the swift and precise hu-man action recognition. The three-dimensional scale invariant feature transform ( 3D SIFT ) descriptor of space-time interest points is utilized to achieve raw feature data by representing the motion video sequences. In order to select the most discriminative samples,AdaBoost algorithm is applied to the raw feature data. The nearest neighbor classifier algorithm is used to test the proposed method on the KTH database. Experiment results show that the method not only achieves the better recognition rate,but also greatly improves the speed of recognition. The validity of the proposed method is proved.