欢迎访问《图学学报》 分享到：
图学学报

图学学报

• 视觉与图像 • 上一篇    下一篇

带有关节权重的DTW 动作识别算法研究

  

  1. 1. 中国农业大学信息与电气工程学院，北京 100083；2. 北京九艺同兴科技有限公司，北京 100083
  • 出版日期:2016-08-31 发布日期:2016-08-09
  • 基金资助:
    国家科技支撑计划项目(2013BAH48F02)

Research on DTW Action Recognition Algorithm with Joint Weighting

  1. 1. College of Information and Electrical Engineering, China Agricultural University, Beijing 100083, China;
    2. Beijing Jiu Yi Tong Xing Technology Co., Ltd, Beijing 100083, China
  • Online:2016-08-31 Published:2016-08-09

PDF (PC)

372

被引次数

11

摘要： 大多数动作仅包含部分关节的运动，现有方法未对运动剧烈的关节与几乎不参与运
动的关节进行区分，一定程度上降低了动作识别精度。针对这个问题，提出一种自适应关节权重
计算方法。结合动态时间规整(DTW)方法，利用获得的关节权重进行动作识别。首先对分类动作
序列进行分段，每段动作序列中运动较剧烈的关节选择分配更高权重，其余关节平均分配权重；
然后提取特征向量，计算两段动作序列的DTW 距离；最后采用K 近邻方法进行动作识别。实验
结果表明，该算法的总体分类识别准确率较高，且对于较相似的动作也能获得较好的识别结果。

关键词: 动作识别, 人体运动分析, 动态时间规整, 关节权重, 姿态特征

Abstract: Human motions always contain only motions of some body parts, but much of the existing
methods on action recognition don’t take the motion intensity of each joint into account, which lower
the accuracy of action recognition in some extent. To solve this problem, an adaptive joint weighting
scheme is proposed to calculate the weight of each joint and combined the weights with dynamic time
warping (DTW) to recognize actions. Firstly, the action sequence was segmented into several
segments and some most violent joints in each segment are assigned higher weight while the
remaining joints are evenly weighted. Then feature vectors of two action sequences were extracted and
the distance between two action sequences were calculated by DTW. Finally the action recognition was
achieved by K-nearest neighbor method. The experiments showed that the overall classification
accuracy of the proposed method is higher, and the result is also good for some similar actions.

Key words: action recognition, human motion analysis, dynamic time warping, joint weight, pose
feature