【作者】 陈勇；

【导师】 黄联芬；

【作者基本信息】 厦门大学 ， 电子与通信工程， 2022， 硕士

【摘要】 目前人口老龄化的趋势越来越明显,独居老年人群体数量日益增大,老年人的照护问题日益凸显。近几年人工智能、物联网的快速发展,推动了我国数字中国的建设和智慧照护领域等发展。人体行为识别是智慧照护领域的研究热点之一。毫米波雷达具有隐私性好、穿透性强、非接触式等独特的优势,能够广泛应用于家居(卧室、浴室)、养老院等场景下的行为检测与识别。因此,对于毫米波雷达的人体行为检测与识别的研究具有重要的实际应用价值。本文的主要工作如下:首先,针对毫米波雷达人体行为检测与识别尚未存在较完善公开数据集的问题,搭建毫米波雷达数据采集系统以及构建毫米波雷达人体行为识别数据集。为了提高毫米波雷达数据标注的可靠性与便捷性,通过设计多线程数据同步采集系统实现了毫米波雷达数据与视频数据的帧同步采集,视频数据作为毫米波雷达数据标注的参照,以解决毫米波数据无法直接通过人眼观察的问题。其次,针对毫米波数据标注需要耗费大量的时间与人力成本,提出并实现了一种毫米波雷达数据自动标注系统。提出了一种基于TCN的毫米波雷达自动标注算法,对包含多个动作的雷达序列数据进行分段、识别与标注。该算法在包含17个不同人的自建数据集中,对4种人体行为动作进行标注的准确率达到了 94.45%。此基础上,为了提高数据集质量,通过上述帧同步采集视频数据对毫米波雷达数据自动标注结果进行进一步检查与纠正。最后,为了充分利用毫米波雷达的频域、时域、空域信息。提出了一种基于多域融合的毫米波雷达人体行为检测与识别算法。针对雷达距离多普勒谱图设计了一种基于注意力机制的卷积神经网络频域特征提取支路;针对人体空间位置信息设计了基于多尺度卷积块的空域特征提取支路,采用中间融合的方法,通过BiLSTM构建时域融合决策网络。实验结果表明,在包含17个人6种不同行为的自建数据集上,基于多域融合的人体行为算法的行为识别准确率达到了 98.16%,比基于距离多普勒人体行为识别算法(97.43%)提高0.73%,比基于空间位置的人体行为算法(74.98%)提高23.18%。今后毫米波雷达行为检测与识别可以从多目标与结合语义信息的复杂人体行为检测与识别方向进一步发展。更多还原

【Abstract】 At present,the trend of population aging is becoming more and more obvious,the number of elderly people living alone is increasing,and the problem of elderly care is becoming more and more prominent.In recent years,the rapid development of artificial intelligence and the Internet of Things has promoted the construction of my country’s digital China and the development of smart care fields.Human behavior recognition is one of the research hotspots in the field of smart care.Millieter-wave radar has unique advantages such as good concealment,strong penetration,and non-contact,and can be widely used in behavior detection and recognition in home(bedroom,bathroom),nursing homes and other scenarios.Therefore,the research on human behavior detection and recognition of millimeter-wave radar has important practical application value.The main work of this paper is as follows:First,aiming at the problem that there is no public data set for millimeter-wave radar human behavior recognition,a millimeter-wave radar data acquisition system and a millimeter-wave radar human behavior detection and recognition data set are constructed.In order to improve the reliability and convenience of millimeter-wave radar data annotation,the frame synchronization acquisition of millimeter-wave radar data and video data is realized by designing a multi-threaded data synchronous acquisition system.The problem that wave data cannot be directly observed by the human eye.Secondly,since millimeter-wave data labeling requires a lot of time and labor costs,this paper proposes and implements an automatic labeling system for millimeter-wave radar data.Firstly,a TCN-based millimeter-wave radar automatic labeling algorithm is proposed to segment,identify and label radar sequence data containing multiple actions.In a self-built dataset containing 17 different people,the algorithm achieved an accuracy of 94.45%for labeling four human actions.On this basis,in order to improve the quality of the data set,the above-mentioned visual aid method is used to further check and correct the automatic annotation results of millimeter-wave radar data by collecting video data in frame synchronization.Finally,in order to make full use of the frequency domain,time domain,and air domain information of millimeter-wave mines.In this paper,a multi-domain fusion-based millimeter-wave radar human behavior detection and recognition algorithm is proposed.A frequency domain feature extraction branch of convolutional neural network based on attention mechanism is designed for range Doppler spectrogram;for human spatial position information,a spatial feature extraction branch based on multi-scale convolution block is designed,using intermediate fusion method to build a temporal fusion decision network through Bi-LSTM.The experimental results show that on the self-built data set containing 6 different actions of 17 people,the human behavior algorithm based on multidomain fusion and using five-fold cross-validation,the average accuracy rate reaches 98.16%,which is higher than that based on distance Doppler human body.The behavior recognition algorithm(97.43%)increased by 0.73%,which was 23.18%higher than the human behavior algorithm based on spatial position(74.98%).In the future,millimeter-wave radar can be further developed in the direction of multi-target and complex human behavior detection and recognition combined with semantic information.更多还原