基于KINECT的静态手势和上肢连续动作识别

发布时间：2018-01-12 16:35

本文关键词：基于KINECT的静态手势和上肢连续动作识别　出处：《北京理工大学》2016年硕士论文　论文类型：学位论文

更多相关文章： 静态手势识别 连续动作识别 KINECT

【摘要】：随着电脑技术的发展,自然简洁的非接触式交互越来越受到人们的青睐。手势和动作作为人们日常交流的常用手段也成为了非接触式交互的主要途径。当前基于单目视觉的手势和动作识别容易受到光照、颜色等环境因素的影响,基于双目视觉的识别虽然能够很大程度上避免这些影响,但是计算过程较为复杂。随着深度摄像机如KINECT的出现,在提供彩色信息的同时还能够提供整个场景的三维信息,因而可以极大程度地减轻环境因素的影响,有效提升手势和动作识别的速度和准确度。本文基于体感摄像机提供的深度数据,对手势识别和动作识别进行研究,主要工作内容如下:1)提出了一种基于深度信息和随机森林模型的日常静态手势识别算法。通过KINECT提供的人体骨架数据对场景深度图分割得到手部的二值化图像,然后通过迭代算法求解手部的最大内接圆进而得到掌心的近似位置,同时通过对手部形状描述符的分析获取手指的具体信息,在此基础之上结合对手势特点的分析提取出简单鲁棒的手势特征,最后选取随机森林算法建立手势的机器学习模型,实现静态手势的识别。实验中对九种静态手势的识别效果表明了该方法的有效性。2)提出了一种基于动态规划和隐式马尔可夫模型的上肢连续动作识别方法。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,并通过先验知识将连续动作序列划分为多个子序列;最后利用隐式马尔可夫模型构造代价函数,通过动态规划算法和一个阈值模型为每个子序列计算最优的动作识别标签,最终实现连续动作识别。实验中对20组上肢连续动作序列的识别结果证明了该方法的有效性。
[Abstract]:With the development of computer technology. Natural and concise contactless interaction is becoming more and more popular. Gestures and actions, as the common means of people's daily communication, have become the main way of contactless interaction. At present, gestures and actions based on monocular vision have become the main way of non-contact interaction. Recognition is vulnerable to light. The recognition based on binocular vision can avoid these effects to a great extent, but the calculation process is more complicated. With the appearance of depth camera such as KINECT. While providing color information, it can also provide three-dimensional information of the whole scene, so it can greatly reduce the impact of environmental factors. It can effectively improve the speed and accuracy of gesture and motion recognition. Based on the depth data provided by the somatosensory camera, this paper studies gesture recognition and motion recognition. The main work is as follows: 1). A daily static gesture recognition algorithm based on depth information and stochastic forest model is proposed. The binary hand image is obtained by segmenting the scene depth map from the human skeleton data provided by KINECT. Then the approximate position of the palm is obtained by iterative algorithm, and the specific information of the finger is obtained by analyzing the shape descriptor of the hand. On the basis of this, the simple robust gesture features are extracted by analyzing the gesture characteristics. Finally, the machine learning model of gesture is established by selecting the stochastic forest algorithm. The recognition effect of nine static gestures shows the effectiveness of the method. 2). A method of upper limb continuous motion recognition based on dynamic programming and implicit Markov model is proposed. Firstly, the normalized skeleton is extracted from the depth information of the video to represent the posture of the human body. Then each action of human body is modeled by implicit Markov model, and the continuous action sequence is divided into multiple sub-sequences by prior knowledge. Finally, the cost function is constructed by implicit Markov model, and the optimal action identification label is calculated by dynamic programming algorithm and a threshold model for each sub-sequence. Finally, the recognition of continuous action is realized, and the effectiveness of the method is proved by the recognition results of 20 groups of upper limb continuous action sequences in the experiment.
【学位授予单位】：北京理工大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】