基于多项式拟合的模特走秀动作分类方法

Classification of catwalk based on polynomial fitting

为自动、准确地对模特走秀动作进行评价,提出了一种基于多项式拟合的动作分类方法。该方法首先利用基于局部亲和域的方法进行人体关节点检测,同时为消除相机视角和个体体型差异性,将检测到的关节点通过普氏分析进行数据校准;其次将人体关节点分为脊柱、上肢和下肢三部分,分别从水平和垂直方向进行多项式拟合;再次对得到的多项式系数进行数据降维;最后将降维后的多项式系数作为动作评价的特征,利用SVM分类器实现模特走秀动作分类。实验结果表明:该方法评价准确率为71.9%,初步实现对人体动作的定性评价。该方法为模特走秀的动作评估提供了一种解决方案。

In order to evaluate catwalk show actions in an automatic and accurate way, an action classification method based on polynomial fitting is proposed in this paper. Firstly, the method uses the Part Affinity Fields to detect the human joint point. Data calibration of detected joint points was conducted by Prussian analysis to eliminate the difference between camera perspective and individual body type. Then, the detected human joint points were divided into three parts: spine, upper limb and lower limbs. Polynomial fitting of the data of the joint points was implemented respectively from the horizontal and vertical directions. Data dimensionality reduction was carried out again for polynomial coefficients obtained. Finally, the polynomial coefficient after reducing the dimension was regarded as the feature of the action evaluation, and SVM classifier was used to achieve classification of the model catwalk show actions. The experimental results showed that the evaluation accuracy of this method was 71.9%, and the qualitative evaluation of human body’s action was preliminarily realized. It provides an effective solution for the catwalk action assessment of the models.