基于时空兴趣点的人体动作识别

Human Action Recognition based on Spatio-Temporal Interest Point

人体动作识别在计算机视觉研究和模式识别领域中逐渐成为一个研究热点。提出一种基于Harris-Laplace时空兴趣点结合3D-SIFT描述子,通过Bag-of-feature构建词袋的方法,并应于用人体动作识别。针对传统Harris算法提取出的兴趣点冗余,所以采用Harris-Laplace算法提取时空兴趣点。3D-SIFT描述子能更好地描述视频序列的本质特征,并且比传统的描述子更有效,Bag-of-feature词袋法表征特征,采用改进的K均值(K-Means)聚类算法进行聚类,最后采用多分类支持向量机(SVM)进行一对一、一对多的分类策略并进行比较。在KTH公开运动数据集上进行实验测试,实验结果证明提出的人体动作识别方法的有效性和鲁棒性。

Human action recognition are increasingly attracting much attention from computer vision and pattern recognition researchers.This paper presents a method based on Harris-Laplace algorithm combined with 3 D-SIFT descriptor,and a Bag-of-feature approach is used to represent videos.The Harris-Laplace algorithm is used to extract the spatial and temporal interest points.The 3 D-SIFT descriptor can better describe the essential characteristics of the video sequence and it is more effective than the traditional descriptor.The K-Means approach is used for clustering.Finally,the support vector machine(SVM） is used as the classifier for human action recognition.One-vs-one and one-vs-rest classification strategies are used and the comparison is made.The experiment on the public database KTH proves the effectiveness and robustness of this method.