基于运动及形状原语的人体动作识别

摘要

It-is-necessary-to-represent-and-model-the-motion-and-shape-information-of-action-in-order-to-recognize-the-action.-Considering-the-fact-that-action-consists-of-a-series-of-sub-actions-namely-primitives,-a-new-algorithm-of-primitive-extraction-and-action-representation-is-proposed-based-on-motion-features-and-shape-features.-According-to-motion-nature-of-action,-motion-feature-is-extracted-through-splitting-the-motion-regions-in-each-frame.-Shape-feature-is-also-extracted-via-histogram-of-gradient.-Then-intra-class-representative-and-inter-class-discriminative-primitives-are-extracted-by-hierarchical-clustering-and-represented-by-Gaussian-Mixture-Model.-Finally,-action-recognition-is-executed-based-on-motion-and-shape-primitives-via-sequence-matching.-The-experiments-show-that-the-proposed-algorithm-obtains-higher-recognition-accuracy-than-existing-ones-on-KTH-and-UCF-datasets,-which-are-95.14%-and-90.00%respectively.%为了进行动作识别，需要对动作的运动与形状信息进行特征表示与建模。由于动作可以由一系列原语即子动作构成，提出了一种新的基于运动特征和形状特征的原语提取及动作表示方法。该算法首先从动作的运动特性出发，通过分割每一帧中的运动区域来提取运动特征，同时利用梯度直方图提取形状特征。通过分层聚类的方式提取具有类内代表性及类间判决性的原语，并利用混合高斯模型对原语进行特征表示。最后利用基于运动及形状原语的序列匹配进行动作识别。实验表明，该算法在KTH和UCF数据集上较现有方法取得了更高的识别正确率，分别为95.14%和90.00%。