| 引用本文: | 蒋留兵，吴岷洋，车 俐. FMCW雷达基于光学字符识别的连续动作识别研究[J]. 雷达科学与技术, 2023, 21(1): 74-81.[点击复制] |
| JIANG Liubing, WU Minyang, CHE Li. Continuous Human Motion Recognition Based on OCR[J]. Radar Science and Technology, 2023, 21(1): 74-81.[点击复制] |
|
| |
|
|
| 本文已被：浏览 5227次 下载 556次 |
码上扫一扫！ |
|
|
| FMCW雷达基于光学字符识别的连续动作识别研究 |
|
蒋留兵，吴岷洋，车 俐
|
|
1. 桂林电子科技大学信息与通信学院， 广西桂林 541004;2. 桂林电子科技大学计算机与信息安全学院， 广西桂林 541004
|
|
| 摘要: |
| 传统的基于雷达的人体动作识别主要采用微多普勒原理，对原始数据进行处理，生成微多普勒时频图，然后输入到基于分类的深度学习网络中进行识别，只能对单个动作进行识别。本文提出一种FMCW雷达光学字符识别技术的连续动作识别方法，首先对采集的雷达数据采用RDM（Range?Doppler Map）向速度维投影的方法逐帧获取微多普勒时频图，然后将处理得到的时频图输入一个特别定制的，由卷积神经网络、inception_resnet、最大池化层和Bi?LSTM的网络组成，使用联结主义时间分类（CTC）作为损失函数进行训练的网络。实验结果表明该方法对步行、跑步、蹲下、站起、跳跃这5种动作的识别准确率分别高达96.16%，95.34%，88.49%，89.37%，96.72%。对一个时间窗口内多个动作的识别也取得了不错的效果，时间上的识别准确率整体令人满意。 |
| 关键词: FMCW雷达 连续人体运动识别 微多普勒 深度学习 光学字符识别 |
| DOI：DOI:10.3969/j.issn.1672-2337.2023.01.009 |
| 分类号:TN958.94 |
| 基金项目:国家自然科学基金资助项目（No.61561010）；广西创新驱动发展专项资助（No.桂科 AA21077008）；桂林电子科技大学研究生教育创新计划资助项目（No.2022YXW07，2022YCXS080） |
|
| Continuous Human Motion Recognition Based on OCR |
|
JIANG Liubing, WU Minyang, CHE Li
|
|
1. School of Information and Communication， Guilin University of Electronic Technology, Guilin 541004， China;2. School of Computer and Information Security, Guilin University of Electronic Technology, Guilin 541004, China
|
| Abstract: |
| Traditional radar?based human action recognition mainly uses the micro?Doppler principle to process the original data to generate a micro?Doppler time?frequency map, and then input it into a classification?based deep learning network for recognition, but only a single action can be recognized. This paper proposes a continuous action recognition method for FMCW radar based on optical character recognition （OCR）. First, the collected radar data is projected by RDM （range?Doppler map） to the velocity dimension to obtain the micro?Doppler time?frequency image frame by frame, and then process it. The obtained time?frequency graph is input into a specially customized network consisting of a convolutional neural network, inception_resnet, a maximum pooling layer, and a Bi?LSTM network, which uses connectionist temporal classification （CTC） as the loss function for training. The experimental results show that the recognition accuracy of this method for walking, running, squatting, standing up, and jumping are as high as 96.16%, 95.34%, 88.49%, 89.37% and 96.72%. The recognition of multiple actions in a time window has also achieved good results, and the accuracy of the recognition in time has reached a satisfactory accuracy as a whole. |
| Key words: FMCW radar continuous human motion recognition micro⁃Doppler deep learning optical character recognition （OCR） |
|