基于傅里叶与局部特征结合的人体姿态识别方法研究

Human Behavior Recognition Method Research Based on Fourier Combined with Local Characteristics

为了提高对人体姿态的识别,提出了一种以人的姿态序列图像的轮廓为特征,包括轮廓的外接矩形的宽高比、形状复杂性变化率、离心率以及傅里叶描述子相结合的人体行为识别方法。首先运用自适应的混合高斯背景建模和形态学结合的方法,利用Canny算子进行边缘检测,实现目标人体轮廓的特征提取。然后采用基于质心边缘距傅里叶描述子与k-means聚类算法与SVM分类器结合的方法,对目标人体轮廓的参数建立具有13个特征值的一维的特征向量,并和RBF神经网络的分类效果进行对比。实验表明,SVM进行分类较为准确,且不需要进行多次的迭代训练,速度较快、识别性能也很好,相比于RBF神经网络而言。运用该方法可以让人体行为识别的正确率在91%以上,该方法简单可行。

In order to improve the recognition of human body posture, A method for action recognition is proposed by using the contours of image sequence as representative descriptors of human action in the text, including the outline of the aspect ratio of the external rectangle, shape complexity rate, eccentricity and Fourier descriptor. Firstly, using the method which Gaussian mixture background is combined with the method of morphology, then use Canny operator for edge detection, to achieve the feature extraction of the human contours. Then a method is adopted which is based on centroid-edge moment of Fourier combined with muliti-class SVM that is combined with k-means clustering algo- rithm, the human body contour extraction of the parameters need set up 13 characteristic value of characteristic vector to classify the feature vectors of human contours and identify human behavior. Experimental results show that this method can achieve the correct recognition rate above 91%.