骨架关节点跟踪的人体行为识别方法

Human Action Recognition Based on Skeleton Joints Tracking

基于彩色图像的运动检测和分割方法难以获取完整的人体骨架,并且只能提取关节点的二维坐标信息,而Kinect传感器能通过捕获深度图像来重建完整的人体骨架关节点三维模型,提高了骨架关节点模型的表示精度。本文通过Kinect的骨架跟踪模块对人体的骨架关节点模型进行提取;然后,提出一种坐标转换方法得到人体骨架关节点的三维坐标表示,利用k均值聚类将关节点坐标量化为符号序列;最后,建立离散隐马尔可夫模型来进行人体行为识别。通过自建的数据集进行实验,实验结果表明:本方法能取得94%的识别率。

It is difficult to get the complete skeleton model by motion detection and segmentation method based on color image which can only get skeleton joints with 2-dimensional information. Kinect sensor can build 3-dimensional human skeleton joints model based on the depth image to improve the accuracy of skeleton joints model. In this paper,skeleton joints location information was obtained by the skeleton tracking module of Kinect. A method of coordinate transformation was put forward to get 3-dimensional coordinate representation of human body skeleton joints. Then vector quantization was performed by using k-means clustering method to represent each action as a symbol sequence. Finally,action recognition was achieved by using trained discrete hidden Markov models. Through the experiment of self-built data sets,the results can achieve 94% recognition rate.