一种基于自相关函数特征的行为识别方法

Action recognition method based on features of autocorrelation function

为了增强特征敏感度,提高人体日常行为识别准确率,针对行为识别特征进行了研究,提出一种基于自相关函数特征的人体行为识别方法。首先对预先采集的人体行为数据进行预处理,然后从时域和频域提取特征后计算得到自相关函数特征,同时采取互相关函数的步进式方法在自相关函数上进行降噪操作。分别使用C4.5决策树、K最近邻、支持向量机、朴素贝叶斯四种分类器进行分类。实验结果表明,与选取纯粹的时、频域特征集进行识别分类的模型相比,选用了包含自相关函数特征的特征集构造出来的模型对行为的识别准确率有较大提高。

In order to enhance feature sensitivity and improve recognition accuracy of human daily actions,this paper researched the features of activity recognition,and presented a human action recognition method based on the autocorrelation function features. The method preprocessed the pre-collected human action data firstly,and then calculated to obtain features of the autocorrelation function after extracting features from the time domain and the frequency domain,and at the same time,it performed noise reduction operations on the autocorrelation function by using a step-by-step method of the cross-correlation function. It separately performed classification by using four classifiers including C4. 5 decision tree,K-nearest neighbor（ KNN）,support vector machine（ SVM）,and naive Bayesian. Experimental results show that,compared with the model recognized and classified by selecting pure time domain and frequency domain feature sets,the model constructed by using feature sets including the features of the autocorrelation function greatly improves actions recognition accuracy.