基于Kinect的动作评价方法研究

发布时间：2018-03-30 23:36

本文选题：Kinect　切入点：姿势识别　出处：《沈阳工业大学》2017年硕士论文

【摘要】：在机器视觉和图像处理领域,关于人体姿势的识别已经成为一项重要的课题,并且在人机交互、虚拟现实和智能视频监控等领域得到广泛的应用。然而目前仍有诸多问题没有得到很好的解决,影响了计算机对于人体行为的理解。由于普通摄像机只能获取到二维图像,但二维信息到三维信息的重建会丢失很多重要数据,影响动作识别的精度。尽管科研人员设计了多种图像重构算法,但是仍无法避免光照、纹理遮挡等影响。而Kinect传感器使用一种新的获取图像的方式,它通过一对红外摄像头捕获到带有空间距离的深度图像,并且在深度图像的基础上提取出含有三维坐标信息的骨骼数据流。但是Kinect没有给出姿势识别的高级函数,原因在于人体动作千变万化,很难构建出一套通用的模型进行识别。为了提升动作识别的效果,本文使用Kinect传感器来获取到人体20个骨骼关节点的三维坐标,并且根据人体姿态的特征,以关节点的相对距离和角度序列为特征参数。在静态姿势的识别中,通过特征向量对样本集进行训练,并使用KNN算法作为分类器对姿势进行识别。在动作评价中,在分析了运动特征序列的时间特性以后,采用线性回归的方法对样本曲线进行训练,使用最小二乘法拟合出一条最佳角度曲线作为标准模板,在考虑到曲线之间时间序列长短不一的问题,通过DTW算法对不同长度的关节角度曲线进行匹配,并且通过定义一套公式对动作进行评价,以曲线之间DTW差值作为实验参数,最终将评价方法应用到动作打分的体感游戏中。本文通过实验分析了相对距离和关节点角度作为动作识别特征向量的可行性。选取简氏太极拳其中的4式作为静态姿势识别对象,实验结果证明通过该方法进行姿态识别可以获得较高的识别率。然后又对动作评价的实验数据进行分析,在总结8个角度的DTW差值样本点分布规律之后,定义一套公式对动作进行评价,并设计动作评估系统论证该动作评价方法的合理性。由于评价公式中的基数和因子会随着动作的变换而不断重新计算,增加了该评价方法的复杂度,接下来的工作是完善动作评价公式的各项参数使评价方法具有更优的效率和适应性。
[Abstract]:In the field of machine vision and image processing, recognition of human posture has become an important issue, and in human-computer interaction, Virtual reality and intelligent video surveillance are widely used. However, there are still many problems that have not been solved well, which affect the understanding of human behavior by computer. But the reconstruction of two-dimensional to three-dimensional information can lose a lot of important data and affect the accuracy of motion recognition. Although researchers have designed a variety of image reconstruction algorithms, but still can not avoid lighting, The Kinect sensor uses a new way to capture images, which capture depth images with spatial distances through a pair of infrared cameras. On the basis of the depth image, the skeletal data stream with three-dimensional coordinate information is extracted. But Kinect does not give a high-level function of posture recognition, because the human body's actions vary greatly. It is very difficult to construct a universal model for recognition. In order to improve the effect of motion recognition, we use Kinect sensor to get the three-dimensional coordinates of 20 skeletal joints, and according to the characteristics of human posture, The relative distance and angle sequence of the node are taken as the characteristic parameters. In the recognition of static pose, the sample set is trained by the feature vector, and the posture is recognized by using KNN algorithm as the classifier. After analyzing the time characteristics of motion feature series, the linear regression method is used to train the sample curve, and the least square method is used to fit an optimal angle curve as the standard template. Considering the difference of time series between curves, the joint angle curve of different length is matched by DTW algorithm, and the action is evaluated by defining a set of formulas. The difference of DTW between curves is taken as the experimental parameter. Finally, the evaluation method is applied to the body feeling game of action scoring. The feasibility of using relative distance and the angle of gate node as the feature vectors of action recognition is analyzed experimentally in this paper. Four of the four forms of Taijiquan are selected as static. State posture recognition object, The experimental results show that the attitude recognition rate can be obtained by this method. Then, the experimental data of motion evaluation are analyzed, and the distribution of DTW difference sample points from 8 angles is summarized. A set of formulas is defined to evaluate the action, and a motion evaluation system is designed to demonstrate the rationality of the evaluation method. The complexity of the evaluation method is increased. The next work is to improve the parameters of the action evaluation formula to make the evaluation method more efficient and adaptive.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】