基于区域形状与运动特征的实时行为识别

Real-time Action Recognition Based on Zone Shapes and Motion Features

提出了一种基于推广的Hu不变矩特征的实时行为识别方法。首先,对Hu不变矩进行改进,使其在离散情况下同时具有平移、旋转和比例不变性。然后,结合运动目标的速度将目标行为刻画成结合Hu矩新特征和速度特征的13维特性向量。其中,Hu矩新特征表征了行为的区域形状特性,速度特征反映了行为的运动特性。随后采用预先定义的一些行为作为先验知识样本训练支持向量机,并最后使用其对待检测行为进行分类以达到行为识别的效果。所提方法计算效率高,能够实时检测人体行为。在处理实拍视频数据的实验中,该方法表现出了理想的处理效率以及识别精度。

This paper presented an efficient action recognition method based on Hu moment invariant features. Firstly, the Hu moment invariants were refined to be new features that are translation, rotation and scale invariant. Then an ac- tion was characterized by a 13-dimensional feature vector consisting of both Hu moment features and action speed fea- tures. The Hu moment features represent the Zone shape of the action, and the action speed features exhibit certain mo- tion characteristics. Finally, a support vector machine（SVM）, which is trained using labeled action frames, was applied to classify test sample actions into different categories. The proposed method is performed on real-world videos and a- chieves acceptable recognition rates with desirable computational efficiencies.