基于特征优选和图像相似度的人体动作识别算法

有效提取特征有利于提高后续人体动作识别的准确率。针对人体动作识别时方向梯度直方图(histogram of oriented gradient,HOG)特征维数过高和相似动作不好区分的问题,提出一种基于特征优选和图像相似度的人体动作识别算法。实验对比三种降维方法主成分分析法(principal component analysis,PCA)、PCA+Pearson、PCA+Spearman处理后的动作识别率,证明PCA+Pearson相关系数的降维效果最佳。同时将全局特征八星模型与降维后的局部特征HOG特征组合在一起全面表征人体动作,并计算相邻两帧图像相似度,自适应分配一个判别周期内单帧支持向量机分类结果的统计权值,最后二次分类人体姿态识别结果。在标准数据集KTH上进行实验,该算法识别准确率为94. 5%,较其他方法有所提高,在视频人体动作识别领域有较好应用价值。

Human Action Recognition Algorithm Based on Feature Selection and Image Similarity

Effective extraction of features is conducive to improve the accuracy of the subsequent human action recognition. Aiming at the problem that the HOG feature dimension is too high and the similar actions are hard to distinguish, a human motion recognition algorithm based on feature selection and image similarity is proposed.Compared the recognition rate of three kinds of dimensionality reduction methods PCA, PAC+Pearson and PCA+Spearman, and proves that the PAC+Pearson correlation coefficient has the best dimensionality reduction effect. Meanwhile the global feature eight star model is combined with the local feature HOG after the dimensionality reduction to represent the human action comprehensively. Then, the similarity of two adjacent frames is calculated, and the statistical weight of single frame SVM classification results is allocated adaptively in a discriminant cycle, and the results of human action recognition are classified twice. The recognition accuracy of the algorithm experimenting on the standard data set KTH is 94.5%, which is better than other methods. It has a good application value in the field of video human action recognition.