基于Fisher矢量编码的运动视频自动评分技术  

Automatic scoring and recognition of action in sports video with Fisher vector encoding

作　　者：石念峰[1,2] 张平[3] 王国强[1] Shi Nianfeng;Zhang Ping;Wang Guoqiang(Luoyang Institute of Science&Technology,Luoyang Henan 471023,China;University of Southern Queensland,Toowoomba Queensland 4350,Australia;Henan University of Science&Technology,Luoyang Henan 471023,China)

机构地区：[1]洛阳理工学院,河南洛阳471023 [2]南昆士兰大学 [3]河南科技大学,河南洛阳471023

出　　处：《计算机应用研究》2018年第10期3138-3141,共4页Application Research of Computers

基　　金：河南省科技攻关项目(182102310041;152102210329;162300410265;172102310635)

摘　　要：传统基于计算机视觉特征的人体运动分析和动作评分技术对局部人体运动特征判别性不强,导致对相似人体动作的类内差异不敏感,自动评分准确率低。提出一种局部时空保持的单目运动视频人体动作特征Fisher矢量(FV)编码方法和自动评分技术。首先提取梯度方向直方图(HOG)和光流直方图(HOF)描述运动视频中人体动作姿态和运动特征,实施2归一化和基于主成分分析的数据降维后获得具有判别性的人体动作特征矢量;然后利用时空金字塔方法在FV编码中嵌入时空特征,提高对动作正确性和协调性的判别能力;最后通过建立不同动作分类的线性模型确定动作评分。在健美操动作自动评分数据集上的实验表明,所提算法的敏感性和特异性约为94.4%和71.4%,与专家评分的中位数平均误差为7.0%,适用于在线体育教学和普通运动训练中基于单目运动视频的动作完成质量评价。Traditional technology on automatic scoring and action analysis is insensitive to the difference of the similar actions due to its poor ability to present local features of human motions,which results a lower accuracy of automatic scoring in the sports video.This paper proposed a Fisher vector encoding with spatial-temporal features locally preserving(STLPFV)and a scoring technology to facilitate automatic rating of human action in the monocular sports video.STLPFV employed the histogram of oriented gradients(HOG)and optical flow(HOF)to feature the pose appearance and motion of the human body in the sports video.By applying 2 normalizations and a PCA-based data dimension reduction to the combination of HOG-HOF,the algorithm received discriminant motion features of the human body.STLPFV combined Fisher vector encoding with spatial-temporal pyramid matching in order to improve its ability to evaluate the correctness and coordination of human action.And the automatic scoring technology used linear models of different action categories to score a set of actions.Experiment results suggest that the sensitivity and specificity of this method are about 94.4%and 71.4%,and the average median error between scores assigned by this method and the expert was 7.0%.The proposed method can be used to evaluate actions in the sports video for the online physical education or general sports training.

关 键 词：Fisher矢量 运动视频 时空特征 高斯混合模型 运动评分 动作完成质量 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]