Recognizing human actions from video sequences is an active research area in computer vision. This paper describes an effective approach to generate compact and informative representations for action recognition. We design a new action feature descriptor inspired from Laban Movement Analysis method. An efficient preprocessing step based on view invariant human motion is presented. Our descriptor is applied in four known machine learning methods, Random Decision Forest, Multi-Layer Perceptron and Multi-class Support Vector Machines (One-Against-One and One-Against-All). Our proposed approach has been evaluated on two challenging benchmarks of action recognition, Microsoft Research Cambridge-12 (MSRC-12) and MSR-Action3D. We follow the same experimental settings to make a direct comparison between the four classifiers and to show the robustness of our descriptor vector. Experimental results demonstrate that our approach outperforms the state-of-the-art methods.
展开▼
机译:从视频序列中识别人类行为是计算机视觉领域的活跃研究领域。本文介绍了一种有效的方法来生成紧凑且信息丰富的表示,以进行动作识别。我们设计了一个新的动作特征描述符,其灵感来自拉班运动分析方法。提出了一种基于视野不变的人体运动的有效预处理步骤。我们的描述符被应用于四种已知的机器学习方法中,即随机决策森林,多层感知器和多类支持向量机(一对一和一对全)。我们建议的方法已经在动作识别的两个具有挑战性的基准上进行了评估,Microsoft Research Cambridge-12(MSRC-12)和MSR-Action3D。我们遵循相同的实验设置,对四个分类器进行直接比较,并显示出描述符向量的鲁棒性。实验结果表明,我们的方法优于最新方法。
展开▼