一种基于深度学习的人体交互行为分层识别方法

A Hierarchical Approach Based on Deep Learning for Human Interactive-action Recognition

本文把人体交互行为分解为由简单到复杂的4个层次:姿态、原子动作、复杂动作和交互行为,并提出了一种分层渐进的人体交互行为识别方法.该方法共有3层:第1层通过训练栈式降噪自编码神经网络把原始视频中的人体行为识别为姿态序列;第2层构建原子动作的隐马尔科夫模型(hidden Markov model,HMM),并利用估值定界法识别第1层输出的姿态序列中包含的原子动作;第3层以第2层输出的原子动作序列为输入,采用基于上下文无关文法(contextfree grammar,CFG)的描述方法识别原子动作序列中的复杂动作和交互行为.实验结果表明,该方法能有效地识别人体交互行为.

This paper discusses the recognition of interaction-level human activities with a hierarchical approach.We classify human activities into four categories：pose,atomic action,composite action,and interaction.In the bottom layer,a new pyramidal stacked de- noising auto-encoder is adopted to recognize the poses of person with high accuracy.In the middle layer, the hidden Markov models （HMMs） of atomic actions are built, and evaluation demarcation algorithm is proposed to detect atomic actions and speed up calcu- lations.In the top layer,the context-free grammar （CFG） is used to represent and recognize interactions.In this layer,a new spatial predicate set is proposed and face orientation is introduced to describe activities.We use Kinect to capture activity videos.The experimental result from the dataset shows that the system possesses the ability to recognize human actions accurately.