| 摘 要: 针对毫米波雷达人体姿势识别方法局限于单一机器学习或深度学习的问题，提出一种基于卷积神经网络和支持向量机的网络结构来实现人体姿势识别。该网络结构通过卷积神经网络对人体姿势的距离角度图进行特征提取，将提取的特征输入分类器中，使用支持向量机代替Softmax进行分类。为了提高训练速度并避免过拟合，使用主成分分析对数据进行降维。利用实际测得的一名实验者摆出的六种人体姿势生成的数据进行验证，实验中每个姿势有200 组数据，共有1,200 组数据，将其80%划分为训练集，其余的20%划分为测试集。实验结果表明，该网络结构较单独使用卷积神经网络拟合速度更快并且准确率得到提升，识别准确率达到100%，整体网络结构简单，具有一定的实用价值。 |
| 关键词: 人体姿势识别；调频连续波；距离角度；卷积神经网络；支持向量机 |
|
中图分类号: TP183
文献标识码: A
|
|
| Research on Human Posture Recognition of Frequency-Modulated Continuous Wave Millimeter-Wave Radar based on CNN-SVM |
|
DENG Zefu
|
(School of Mechanical Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China )
203601623@st.usst.edu.cn
|
| Abstract: Aiming at the problem that the human posture recognition method of millimeter-wave radar is limited to single machine learning or deep learning, this paper proposes a network structure based on CNN-SVM (Convolutional Neural Network and Support Vector Machine) to realize human posture recognition. CNN is used in this network structure to extract the distance angle map features of human postures, which are input into a classifier and then classified by SVM instead of Softmax. In order to improve the training speed and avoid overfitting, principal component analysis is used to reduce the dimension of data. Actually measured data generated by an experimenter who presents six human postures are used for verification. In the experiment, there are 200 sets of data for each posture, a total of 1200 sets of data, 80% of which are divided into training sets and the remaining 20% are for testing sets. The experimental result show that the fitting speed of the proposed network is faster and the accuracy is improved, compared with the network structure that uses the convolutional neural network alone, and the recognition accuracy rate reaches 100%. The overall network structure is simple, and has certain practical value. |
| Keywords: human posture recognition; FMCW; distance angle; convolutional neural network; support vector machine |
