融合深度学习特征的人体摔倒检测研究

发布时间：2018-03-14 02:08

本文选题：摔倒检测　切入点：视频监控　出处：《华东师范大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着社会的发展,全球的老龄人口在不断地增加,使老年人的医疗健康成为社会关注的问题之一。在影响老人身体健康的问题中,摔倒伤害所占的比率最高。本论文以深度学习为基础,研究日常生活中的人体摔倒检测问题。首先,通过比较视频序列和图片中的人体检测方法,选择前景提取法进行运动人体的初始检测。然后,融合深度学习框架提取的高维特征以及基于图像的低维特征表示人体的运动状态。对于人体摔倒的判断,构建级联的支持向量机判别模型。第一层支持向量机实现某一时刻人体运动状态的判断,第二层支持向量机使用连续时间内人体的运动状态进行摔倒的最终判断。最后,根据实际生活中可能出现的问题,引入人体描述符与分类器结合的检测方式,通过PCANet和迭代梯度提升树提高了复杂状态下运动人体的检测能力。本文的人体摔倒检测模型在自己收集的数据集上获得了 97%的敏感度以及95%的特异性,在公开的摔倒数据集获得的检测结果为93.7%的敏感度和92.0%的特异性。
[Abstract]:With the development of society, the aging population in the world is increasing, and the medical health of the elderly has become one of the social concerns. Fall injury is the highest rate. Based on deep learning, this paper studies the human fall detection in daily life. Firstly, we compare the human body detection methods in video sequences and pictures. The foreground extraction method is selected for the initial detection of moving human body. Then, the high dimensional feature extracted by the depth learning framework and the low-dimensional feature based on image represent the motion state of the human body. A cascade discriminant model of support vector machine is constructed. The first layer realizes the judgment of the human body's motion state at a certain time, the second layer uses the moving state of the human body during the continuous time to make the final judgment of falling down. Finally, According to the possible problems in real life, the method of combining human descriptors with classifiers is introduced. Through PCANet and iterative gradient lifting tree, the detection ability of moving human body in complex state is improved. In this paper, the sensitivity of 97% and the specificity of 95% are obtained on the data set collected in this paper. The results obtained in the open fall dataset were 93.7% sensitivity and 92.0% specificity.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】