摘要
传统人体动作识别分类器对异常值比较敏感,容易受固有噪声影响,这会导致严重的类失衡问题,所以相似的人体行为可能存在很大类内差异。提出一种基于能量的最小二乘双分界面支持向量机(ELS-TSVM)的人体动作识别算法。ELS-TSVM是LS-TSVM的有效改进,采用两个超平面,每个超平面引进能量参数来减少噪声和异常值的影响。首先对于输入的视频使用梯度方向直方图特征和光流直方图特征识别人体动作;然后检测可能的兴趣点,生成时空特征后提取时空视觉词袋特征,通过构建一组视觉词袋来完成特征提取;最后,利用ELSTSVM完成分类。在Weizmann和Hollywood数据库上的实验验证了该算法的有效性及可靠性,相比其他几种较新方法,该算法更加高效精确,且大大减少了算法执行时间。
Traditional classifier of human action recognition is sensitive to outliers and susceptible to the inherent noise, which results in severe class imbalance problem, so similar human behaviors may exist a wide class differences. This paper proposed the energy-based least square twin support vector machine （ ELS-TSVM ） human behavior recognition classifier. ELS-TSVM was an effective improvement to LS-TSVM, using two hyper-planes, and introduced energy parameter in each hyper-plane to reduce the effects of noise and outliers. Firstly, it used the gradient direction histogram feature （HOG） and optical flow histo- gram features （HOF） to realize human behavior. Then, detected the possible interesting points to generate spatial and tempo- ral characteristics and bag of features （BoFs）. It completed the feature extraction by building a set of visual word of bag. Fi- nally, accomplished the recognition by ELS-TSVM. This paper used Weizmann and Hollywood databases , the experimental results show that the proposed method is more efficient and accurate, and the running time is greatly reduced, compared to several other relatively new methods.
出处
《计算机应用研究》
CSCD
北大核心
2016年第2期598-601,631,共5页
Application Research of Computers
基金
国家自然科学基金资助项目(61103143)
中国博士后科学基金资助项目(2012M512008)
新疆维吾尔自治区自然科学基金项目(2010211A08)
关键词
多分类识别
类失衡
双分界面支持向量机
人体动作识别
最小二乘法
multi-class classification
class imbalance
twin support vector machine
human action recognition
least square method
