基于单目摄像头的智能电视手势交互系统

发布时间：2018-04-22 06:37

本文选题：手势跟踪 + 增量子空间　；参考：《华南理工大学》2014年硕士论文

【摘要】：手势作为一种自然的人机交互方式被广泛应用于智能电视交互中。考虑到手势的非刚性形变和其快速运动的特点，外加家居环境中不确定的复杂的光照条件和背景环境，以及计算复杂度、实时性和低成本的考虑，智能电视手势交互系统的研究依旧面临诸多挑战。本文针对创维42E790U智能电视，基于单目摄像头设计开发一套智能电视手势交互系统。基于计算机视觉的手势识别过程大致可分为手势检测、手势跟踪以及手势识别三个阶段。本文在已有的相关技术基础上，主要做了如下三部分工作：（1）手势跟踪方面：本文的改进基于原有的增量子空间跟踪模型。本文中所采用的跟踪算法可以分为两个模块，分别是运动模型和目标外观模型。运动模型减小了目标位置的搜索范围，并给出一定的运动预测，为跟踪的实时性提供保障。目标外观模型为准确定位目标提供保障。本文应用中，为了适应手势的非刚性形变及家居空间中复杂的光照、背景环境，采用目标外观模型可以自适应更新的增量子空间在线跟踪模型。为了减轻跟踪过程中背景对跟踪造成的影响，采用加权的增量子空间在线跟踪模型。为了适应手势的快速运动，基于粒子滤波框架提出二次采样和运动预测的解决方案。实验结果显示了本文算法的有效性。（2）手势识别方面：本文先对电视系统的交互任务进行分析，设计了一套简单的、易用且交互性较好的动态手势指令集。并为这套指令集设计了对应的动态手势识别算法。本文系统采用基于编辑距离的识别方法，通过对轨迹的预处理、有效性检验和基于标准差改进的轨迹相似度分类，本文的动态手势系统识别准确率达到了96%。取得非常好的效果。（3）系统设计方面：本文基于电视资源的分析，将算法资源进行有效分配，将复杂的算法运算置于高配置的PC端处理，而对指令的响应和显示则由智能电视系统来处理，两者通过网络进行通信，以达到系统的高效协同运作。实验结果表明，本系统设计方法具有一定的实用价值。相比于智能电视的内置手势控制系统，，有更好的兼容性和拓展性。
[Abstract]:As a natural human-computer interaction, gesture is widely used in intelligent TV interaction. Taking into account the non-rigid deformation of gestures and the characteristics of their rapid motion, the uncertain and complex lighting and background conditions in additional home environments, as well as the considerations of computational complexity, real-time performance and low cost, The research of intelligent TV gesture interaction system still faces many challenges. In this paper, an intelligent TV gesture interaction system based on single camera is designed for Skyworth 42E790U intelligent TV. The process of gesture recognition based on computer vision can be divided into three stages: gesture detection, gesture tracking and gesture recognition. Based on the existing related technologies, this paper mainly does the following three parts of work: 1) gesture tracking: the improvement of this paper is based on the original incremental subspace tracking model. The tracking algorithm used in this paper can be divided into two modules: motion model and object appearance model. The motion model reduces the search range of the target position and gives a certain motion prediction, which provides a guarantee for the real-time tracking. The object appearance model provides the guarantee for the accurate positioning of the target. In order to adapt to the non-rigid deformation of gestures and the complex illumination and background environment in home space, the incremental subspace online tracking model can be adaptively updated by the target appearance model. In order to reduce the influence of background on tracking, a weighted incremental subspace online tracking model is adopted. In order to adapt to the fast motion of gestures, a solution of quadratic sampling and motion prediction is proposed based on particle filter framework. Experimental results show the effectiveness of the proposed algorithm. (2) gesture recognition: firstly, the interactive task of TV system is analyzed, and a set of simple, easy-to-use and interactive dynamic gesture instruction set is designed. The corresponding dynamic gesture recognition algorithm is designed for the instruction set. In this paper, an edit-distance based recognition method is adopted, and the accuracy of dynamic gesture recognition is 96% through the pre-processing of trajectory, validity test and improved trajectory similarity classification based on standard deviation. Very good results have been achieved. 3) system design: based on the analysis of TV resources, this paper allocates the algorithm resource effectively, and puts the complex algorithm operation on the high configuration PC terminal, while the response and display to the instruction is handled by the intelligent TV system. In order to achieve the efficient and cooperative operation of the system, the two systems communicate through the network. The experimental results show that the system design method has certain practical value. Compared with intelligent TV's built-in gesture control system, it has better compatibility and expansibility.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN949.2

【相似文献】