RGB-D视频中连续动作识别

Recognizing continuous human actions from RGB-D videos

提出了一种基于动态规划的方法识别深度视频中的动作序列,首先通过视频中的深度信息提取出归一化的骨架来表征人体的姿态;然后通过隐式马尔可夫模型对人体的每个动作进行建模,并利用先验知识对整个连续动作序列进行划分;最后利用隐式马尔可夫模型输出概率构造代价函数,通过动态规划求解得到最优的动作识别标签,实现连续动作识别。实验结果验证了该方法对连续动作识别的高效性和准确性。

A novel algorithm for continuous actions recognition from RBG-D videos is proposed in this paper.Firstly,at each frame,the human body is represented by normalized skeleton extracted from the depth cues.Then,we employ hidden Markov model to build a model for each action,and segment the entire continuous actions sequence based on priori knowledge.Finally,by minimizing the cost function constructed by HMM probabilities,an optimal action label is computed with the help of dynamic programming.The experiments demonstrate the efficiency and accuracy of our algorithm on recognizing continuous actions.