基于加速度传感器的人体行为识别研究

发布时间：2018-06-01 16:10

本文选题：加速度传感器 + 人体多样性　；参考：《天津工业大学》2017年硕士论文

【摘要】：近年来,随着微电子和传感器技术的不断进步以及模式识别理论的深入研究,基于加速度传感器的人体行为识别得到快速发展。随着社会的进步,人们对智能交互和健康监护等方面的需求也日益增长,所以基于加速度传感器的人体行为识别在医疗保健、运动检测、能耗评估等领域受到了广泛的关注和研究。与基于加速度传感器的行为识别相对应的是基于计算机视觉的行为识别,相比之下,基于加速度传感器的行为识别更能体现人体运动的本质,并且不受场景或者时间的限制,携带方便,成本较低,更适合推广应用。目前,在人体行为识别领域中,分类模型通常有两种,分别是通用模型和个性化模型。但是通用模型没有考虑人体多样性问题,不能适用于所有人,而个性化模型需要人为干预较多,本文针对这两种模型的不足提出了一种折衷的模型训练方法,即对人体多样性因素分区间后的原始加速度数据进行训练得到多个模型。另外,为了使识别模型适用于更加广泛的情况,在数据采集阶段还考虑了加速度传感器的位置。该方法使得模型更具普遍性的同时又能够提高识别精确度,通过对人体静止、走路、跑步、上下楼梯五种行为进行试验,识别率达到了 95%左右。实验表明该方法是切实有效的。采集多种多样的人的加速度数据并非易事,所以笔者又从算法角度进行研究来提高行为识别的精确度。SVM算法最初是用来解决二分类问题的,对于多分类问题的效果并不是很好。为了使SVM算法更好地解决多分类问题,本文使用基于二叉树的SVM,整合了 SVM和二叉树的特点,并使用最短距离算法进行聚类使得分类更加准确。由于SVM的分类精度主要取决于核函数与参数的选取,所以本文又结合了网格搜索算法选取核函数的惩罚系数c和核参数γ。最终实验识别率达到了 90%以上,通过与其他算法的结果对比,证明该方法也是可行的。
[Abstract]:In recent years, with the continuous progress of microelectronics and sensor technology and the in-depth research of pattern recognition theory, the acceleration sensor based human behavior recognition has been rapidly developed. With the progress of society, the demand for intelligent interaction and health monitoring is also increasing, so the acceleration sensor based human behavior recognition in health care, motion detection, Energy consumption assessment and other fields have received extensive attention and research. The behavior recognition based on acceleration sensor is based on computer vision. In contrast, the behavior recognition based on acceleration sensor can reflect the essence of human motion and is not limited by scene or time. Easy to carry, low cost, more suitable for popularization and application. At present, in the field of human behavior recognition, there are usually two kinds of classification models, one is general model and the other is personalization model. But the universal model does not consider the diversity of human body and can not be applied to everyone, but the personalized model needs more human intervention. In this paper, a compromise model training method is proposed to overcome the shortcomings of the two models. In other words, the original acceleration data of human diversity factors are trained to obtain multiple models. In addition, in order to make the recognition model suitable for a wider range of cases, the position of the acceleration sensor is also considered in the data acquisition stage. This method not only makes the model more universal but also can improve the recognition accuracy. The recognition rate is about 95% through the experiments on five behaviors of human body such as stillness walking running and going up and down stairs. Experiments show that the method is effective and effective. It is not easy to collect the acceleration data of a variety of people. Therefore, the author studied from the point of view of algorithm to improve the accuracy of behavior recognition. SVM algorithm was initially used to solve the two-classification problem, but the effect of multi-classification problem is not very good. In order to solve the multi-classification problem better, this paper uses SVM based on binary tree, integrates the characteristics of SVM and binary tree, and uses the shortest distance clustering algorithm to make the classification more accurate. Since the classification accuracy of SVM mainly depends on the selection of kernel functions and parameters, this paper combines the grid search algorithm to select the penalty coefficient c of kernel function and the kernel parameter 纬. Finally, the experimental recognition rate is over 90%, and it is proved that this method is feasible by comparing with the results of other algorithms.
【学位授予单位】：天津工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.4;TP212

【相似文献】