融合时空兴趣点和多元广义高斯混合模型的人体动作识别

Human Motion Recognition based on Spatiotemporal Interest Points and Multivariate Generalized Gaussian Mixture Models

人体动作识别近年作为计算机视觉领域的热点研究方向,被广泛用于人机交互、虚拟现实等领域。针对传统人体动作识别算法中提取特征时冗余点过多、忽略图像数据的关联性等问题,提出一种融合时空兴趣点和结合定点估计的多元广义高斯混合模型(MGGMMs)的人体动作识别方法,通过过滤冗余特征点和利用多元广义高斯混合模型实现了特征点的有效提取以及对数据关联性的充分利用。以改进的Harris-Laplace算法和3D-SIFT描述子提取视频序列的特征点,利用BOW模型进行视觉词聚类,最后通过改进的多元广义高斯混合模型进行建模和分类。在KTH公开数据集上进行实验,实验结果表明提出的人体动作识别方法能够对视频中人体动作进行有效识别和分类。

Human action recognition has been widely used in the fields of human-computer interaction and virtual reality, etc. Aiming at the problems of excessive redundancies and neglecting the correlation of image data in traditional human motion recognition algorithm, this paper proposed a human motion recognition method based on spatio-temporal interest points and multivariate generalized Gaussian mixture model combined with fixed-point estimation. Redundant feature points and multivariate generalized Gaussian mixture models are used to effectively extract feature points and make full use of data correlation. The Harris-Laplace algorithm and 3D-SIFT descriptor extracted the feature points of the video sequence, and the BOW model clustered the visual words. Finally, the improved multivariate generalized Gaussian mixture model modeled and classified. Experiments were performed on the KTH datasethe experimental results show that the proposed human motion recognition method can effectively recognize and classify human motion in video.