Abstract: Aiming at the problem of timeliness in human-computer interaction using human action based on Kinect, a rapid human action recognition method based on time series similarity is proposed. Three-dimensional coordinates of the key points extracted from 20 joints of human body obtained from the Kinect sensor are transformed into eigenvector and the eigenvector model?is a superb action descriptor. A Fast Dynamic Time Warping（F-DTW） algorithm is proposed to solve the speed problem of action recognition. Lower bound function and early termination technique is used to solve the time-lag problem so as to control the robot quickly. 20 kinds of actions are defined for recognition, the average recognition rate is greatly improved than traditional algorithm, verifying the validity of the method and meeting the requirements of the interaction with robot.

Key words: action recognition, time series similarity, Fast Dynamic Time Warping（F-DTW） algorithm, Kinect, robot

摘要： 针对在Kinect平台利用人体动作进行人机交互的时效性问题，提出了一种基于时间序列相似性的快速人体动作识别方法。通过Kinect获取人体全身20个关节点，提取关键点的空间三维坐标，转化成特征向量，该特征向量模型能很好地对全身动作进行表示；在动作识别方面提出了一种快速动态时间弯曲距离（Fast Dynamic Time Warping，F-DTW）算法，解决了因动作速度不同导致的两时间序列在时间轴上不一致的问题，通过引入下界函数和提前终止技术对算法进行加速优化，解决动作识别的时延问题，从而能快速地控制机器人；定义20种动作进行识别，平均识别速度较传统算法大大提高，验证了方法的有效性，满足与机器人交互的要求。

关键词: 动作识别, 时间序列相似性, F-DTW算法, Kinect, 机器人