基于Android平台的跌倒检测算法研究及实现

发布时间：2018-02-20 05:47

本文关键词： 跌倒检测支持向量机加速度角速度多特征融合　出处：《杭州电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着全球人口老龄化程度的持续加剧,人体跌倒检测、预警与防护等技术,成为预防老人意外跌倒伤害和提升生活品质的重要保证。智能手机功能的不断增强,可穿戴式计算与人体传感器网络技术的不断成熟,利用这些设备通过科学的手段识别老年人跌倒成为国内外研究的新热点。这一技术不仅在理论研究方面具有较高的意义,而且在实际应用方面的价值也非常大。跌倒识别问题的研究主要分为四大部分,跌倒模型的建立、实验方案设计、原始数据的处理及判别算法的应用。选择人体运动过程中的加速度和角速度作为主要的特征量,结合阈值法、支持向量机、决策树等算法实现跌倒检测。跌倒检测实质上是简单的二分类问题,主要目标是区分一个活动属于跌倒还是日常行为活动。阈值法思想比较直观,算法实现相对容易,是最普遍的研究方法。由于目前尚无确定的模型,阈值的大小大多基于经验判断或统计实验结果得出,它的不足是阈值的设定对识别结果影响很大,稳定性较差。模式识别的方法针对大样本分类有很大的优势,然而在实际研究中,难以获取大量真实跌倒数据,大都基于小样本的研究。支持向量机对于解决较小样本和非线性模式识别等问题都具有一定的优势。本文致力于通过基于支持向量机的运动特征加权方法实现跌倒检测。首先,论文分析智能手机传感器的输出数据,选择并提取了区分跌倒与日常行为活动的特征量,并针对样本量少的问题,在特征样本基础上利用高斯分布原则构造了跌倒与日常行为活动的虚拟样本;其次,改进特征融合的方式,采用特征加权核函数实现加速度及角速度特征的加权融合,构建基于支持向量机的跌倒识别模型;然后,对跌倒识别模型关键部分进行分析,验证算法泛化性能。最后在安卓手机平台上实现了以本文所构建的识别模型为核心的跌倒检测系统,对跌倒检测算法的实际应用效果进行了验证。实验结果表明:本文构建的特征融合的跌倒监测算法在实际应用中,敏感度为90.67%,特异度为92.96%,准确度为92.14%,验证了本文构建的检测算法的检测性能,同时也侧面验证了基于安卓平台的跌掉检测软件的健壮性和稳定性。
[Abstract]:As the aging of the global population continues to increase, human fall detection, early warning and protection technologies have become an important guarantee to prevent accidental fall and injury and improve the quality of life of the elderly. With the development of wearable computing and human sensor network technology, it has become a new research hotspot at home and abroad to use these devices to identify the fall of the elderly by scientific means. The research on fall recognition is mainly divided into four parts: the establishment of fall model, the design of experimental scheme, The processing of raw data and the application of discriminant algorithm. The acceleration and angular velocity in the process of human body motion are selected as the main characteristic quantities, combining with the threshold method, the support vector machine, In fact, fall detection is a simple two-classification problem. The main goal is to distinguish one activity from the other. The threshold method is more intuitive, and the algorithm is relatively easy to implement. Because there is no established model at present, the threshold value is mostly based on empirical judgment or statistical experimental results. Its disadvantage is that the threshold setting has a great influence on the recognition results. The method of pattern recognition has a great advantage in classifying large samples, but it is difficult to obtain a large number of real fall data in practical research. Most of the researches are based on small samples. Support vector machines (SVM) have some advantages in solving the problems of small samples and nonlinear pattern recognition. In this paper, the motion feature weighting method based on support vector machine (SVM) is used to realize fall detection. This paper analyzes the output data of smart phone sensor, selects and extracts the characteristic quantity to distinguish fall from daily behavior, and aims at the problem of small sample size. Based on the feature sample, the virtual sample of fall and daily behavior is constructed by using Gao Si's distribution principle. Secondly, the method of feature fusion is improved, and the weighted fusion of acceleration and angular velocity is realized by using feature weighted kernel function. The fall recognition model based on support vector machine (SVM) is constructed, and then the key parts of the fall recognition model are analyzed. Finally, a fall detection system based on the recognition model built in this paper is implemented on the Android mobile platform. The experimental results show that the feature fusion algorithm proposed in this paper is applied in practice. The sensitivity is 90.67, the specificity is 92.96 and the accuracy is 92.14. It verifies the detection performance of the detection algorithm constructed in this paper, and also verifies the robustness and stability of the drop detection software based on Android platform.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212.9;TP316

【参考文献】