基于递归神经网络的人体行为识别方法研究与系统实现

发布时间：2018-03-20 16:27

本文选题：行为识别　切入点：递归神经网络　出处：《哈尔滨工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：目前,大多数人体行为识别问题的研究聚集在视频数据上,而基于视频数据的行为识别方法容易侵犯个人隐私,同时由于视频图像数据的复杂性造成识别准确度不高。随着物联网技术的快速发展,使用穿戴设备来识别人体行为逐渐成为人体行为识别的一种新办法,吸引了一大批专家学者的目光。研究人员尝试使用了许多机器学习的办法,如:随机森林、支持向量机等浅层学习方法,在实验室环境下表现不错,但是距离实际应用还有一段距离。由于人体行为属于一种连续型的时间序列数据,同时考虑到递归神经网络在时间序列数据上表现出来的神奇效果,以及在大量实际应用中均取得不错的结果,因此本文使用深层学习中的递归神经网络模型来解决行为识别问题。本文具体的研究工作和创新点包括:1.本文基于穿戴设备,依据人体行为的特点,提出了一种先识别原子活动,后识别复杂活动的分层级识别行为的方法。2.本文将原子活动定义为日常活动所包含的短周期活动和外出交通方式活动,包括跌倒、下楼、上楼、电梯向下、电梯向上、跑、坐、站、走、坐地铁、坐公交车、坐小汽车、躺共计13种活动。使用面向原子活动识别的递归神经网络模型构建识别算法,同时进行了大量的实验,在实验室环境和实际应用环境均取得了较好的结果。3.本文将复杂活动定义为中长期的生活活动,包括但不限于做饭、拖地、扫地、吃饭、看电视等。本文参考时间序列分析中模式匹配的思想设计了基于动态弯曲距离的复杂活动识别算法。该算法在实验室环境下,能够识别睡觉、吃饭、锻炼三种复杂活动,具有较低的平均索引误差(复杂活动识别算法的评价指标,详情见4.2章)。4.本文实现了一个用以完成数据采集、数据分析、数据展示、数据管理的行为识别原型系统以及应用本文研究成果的智慧养老系统。
[Abstract]:At present, most researches on human behavior recognition are focused on video data, and behavior recognition methods based on video data are apt to violate personal privacy. At the same time, because of the complexity of video image data, recognition accuracy is not high. With the rapid development of Internet of things technology, the use of wearable devices to identify human behavior has gradually become a new method of human behavior recognition. The researchers tried to use a lot of machine learning methods, such as random forest, support vector machine and other shallow learning methods, and performed well in the laboratory environment. But there is still a long way to go. Because human behavior is a continuous type of time series data, and considering the magical effect of recurrent neural network on time series data, And in a large number of practical applications have achieved good results, so this paper uses the recursive neural network model in deep learning to solve the behavior recognition problem. The specific research work and innovation of this paper include: 1.This paper is based on wearable devices. According to the characteristics of human behavior, this paper proposes a hierarchical method of recognizing atomic activity first and then complex activity. In this paper, atomic activity is defined as short period activity and outgoing traffic mode activity included in daily activities. Including falling down, going downstairs, going upstairs, elevators down, elevators up, running, sitting, standing, walking, taking the subway, taking buses, taking cars, lying down a total of 13 activities. At the same time, a large number of experiments have been carried out, and good results have been obtained in both the laboratory environment and the practical application environment. In this paper, complex activities are defined as medium- and long-term life activities, including, but not limited to, cooking, mopping, sweeping, eating. Based on the idea of pattern matching in time series analysis, a complex activity recognition algorithm based on dynamic bending distance is designed in this paper. The algorithm can recognize three complex activities, sleep, eat and exercise in laboratory environment. It has low average index error (evaluation index of complex activity recognition algorithm, see chapter 4.2, chapter 4. 4. This paper implements a method to complete data acquisition, data analysis, data display, etc. The behavior recognition prototype system of data management and the intelligent pension system applying the research results of this paper.
【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;TP183

【相似文献】