摘要
为了能够在丰富复杂的网络信息中快速找到所需图片,提出一种基于骨架特征的人体上半身动作分类方法,以提高相应图片的检索效率。对人体运动图片进行人体运动时上半身姿势识别,得到能够表示人体位置、方向以及大小的"火柴人模型"(即骨架特征),使用矩阵形式对提取到的骨架特征进行描述。为了校正因距离和位置变化造成的尺度差异,对特征矩阵进行归一化处理,然后使用多分类SVM方法对提取的骨架特征进行训练,得到可以对不同动作进行分类的分类器。以收集到的人体运动图片作为测试数据库进行实验,实验结果表明,该算法的分类准确率达到97.36%,能够很好地对人体动作进行分类。同时,在Buffy数据库上进行图片检索对比实验,实验结果表明,所提算法的分类准确率更高,更好地提高了图片检索效率。
In order to find the desired pictures quickly in the abundant and complex network information, a method for human upper-body action classification based on skeleton features is proposed to improve the efficiency of the corresponding pictures. It does the pose estima- tion for the image of human motion, acquires the ＂stickman＂ （ skeleton features） representation of the location, orientation, and size of body parts, and describes the skeleton features with matrix form. In order to correct the scale differences caused by distance and position changes, the feature matrix is normalized. Then the multi-classification SVM is used to train the skeleton features and obtain the classifier which can classify different actions. The images of human motion collected are as the test data for experiments which show that its classi- fication accuracy reaches 97.78% and it can do well in human action classification. At the same time,an image retrieval contrast experi- ment is done on the Buffy database, which show that it has higher classification accuracy and enhance image retrieval efficiency better.
出处
《计算机技术与发展》
2017年第8期83-87,共5页
Computer Technology and Development
基金
江苏省自然科学基金(BK20130883)
南京邮电大学引进人才科研启动基金(NY212016
NY214189)
关键词
动作分类
姿势识别
骨架特征
多分类SVM
action classification
pose estimation
skeleton features
multi-class SVM
