基于隐性马尔科夫模型的分组手势识别

发布时间：2018-05-12 07:36

本文选题：手势分组 + 隐性马尔科夫　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：使用可穿戴式设备进行手势识别正在快速地成为一个研究热点,被广泛地应用在了行为检测、手语识别以及人机交互方面。今天,随着微机电系统(Micro Electromechanical System,MEMS)的发展使得生产更小、更轻便的传感器和设备成为了可能,这些设备可以穿戴在人身上用来检测人的行为甚至一些更小的肢体动作。当用户在资源受限的穿戴式设备上执行实际的手势识别的时候,需要考虑识别精度和算法的时间复杂度。目前许多的手势识别算法已经被提出并采用了。隐性马尔科夫模型(Hidden Markov Model,HMM)是目前用的比较多的手势识别算法,HMM最早多被用于语音识别,由于手势序列和语音序列的相似性,因此被广泛用于手势识别当中,且可以达到比较高的识别精度。由于HMM的高运算复杂度,导致当用户在资源受限的移动设备上使用HMM来进行手势识别的时候,达不到实时反应的效果,用户体验差,需要进行改进。HMM的运算复杂度直接与需要识别的数据集规模、观察序列的长度以及状态数目成正比,减少这三个参数的值能够降低运算复杂度,但是识别精度也相应的变低了。因此有必要找出一种方法:在维持识别精度的前提下尽量减低算法运算复杂度,使其适应移动终端的运算能力。针对上述提出的问题,本文提出了一种通过对手势进行分组来降低识别算法的运算复杂度以及通过为每个组设定不同的HMM来保持甚至提高识别精度的方法。此方法包含三部分内容:手势分组,组模型的建立,以及每个组里面手势模型的建立。手势分组使用基于K-means++的方法;组模型使用基于表格的方法;手势模型使用HMM,同一个组内的HMM具有相似的结构,不同组具有不同的结构。为了验证分组手势识别方法的有效性,本文定义了12种手势,这些手势考虑到了不同的形状、方向以及重复性,具有代表性。然后,本研究搭建了一个数据采集平台,这个平台包括可穿戴式的硬件平台以及电脑端上位机平台。通过这个平台,本研究采集了大量的手势数据来验证分组手势识别方法的有效性。实验结果表明:在不损失识别精度的前提下,和标准的HMM相比,本文提出的对手势进行分组手势识别的方法的运算复杂度大大降低,验证了方法的可行性。
[Abstract]:Gesture recognition using wearable devices is becoming a research hotspot and has been widely used in behavior detection, sign language recognition and human-computer interaction. Today, with the development of MEMS Micro Electromechanical Systems, it is possible to produce smaller, lighter sensors and devices that can be worn on people to detect human behavior and even smaller body movements. When the user performs actual gesture recognition on a wearable device with limited resources, the recognition accuracy and the time complexity of the algorithm should be considered. At present, many gesture recognition algorithms have been proposed and adopted. Hidden Markov Model HMMM (Hidden Markov Model) is one of the most widely used gesture recognition algorithms at present. Due to the similarity between gesture sequences and speech sequences, hmm is widely used in gesture recognition. And it can achieve higher recognition accuracy. Due to the high computational complexity of HMM, when users use HMM for gesture recognition on resource-constrained mobile devices, the real-time response is not achieved and the user experience is poor. The computational complexity of the improved. Hmm is directly proportional to the size of the dataset to be identified, the length and the number of states of the observation sequence. Reducing the values of these three parameters can reduce the computational complexity, but the recognition accuracy is also reduced accordingly. Therefore, it is necessary to find a way to reduce the computational complexity of the algorithm under the premise of maintaining the recognition accuracy, so that it can adapt to the computing ability of the mobile terminal. To solve the above problems, this paper proposes a method to reduce the computational complexity of the recognition algorithm by grouping the gestures and to maintain or even improve the recognition accuracy by setting different HMM for each group. The method consists of three parts: gesture grouping, group modeling, and gesture modeling within each group. Gesture grouping uses K-means based approach, group model uses tabular approach, gesture model uses HMMs, and HMM in the same group has similar structure and different groups have different structures. To verify the effectiveness of the grouping gesture recognition method 12 gestures are defined which take into account different shapes directions and repeatability and are representative. Then, a data acquisition platform is built, which includes wearable hardware platform and PC platform. Through this platform, this study collected a lot of gesture data to verify the effectiveness of the grouping gesture recognition method. The experimental results show that the computational complexity of the proposed method is greatly reduced compared with the standard HMM without loss of recognition accuracy, and the feasibility of the method is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】