| Human head movement recognition has a very broad application prospect in the fields of disability care,driver fatigue monitoring,and intelligent guidance.Human head movement recognition using millimeter wave radar has received increasing attention due to its advantages such as being unaffected by light,strong privacy,and non-contact detection.At present,head movement recognition based on millimeter wave radar has problems such as low recognition accuracy and fewer types of recognition actions.In response to the above issues,based on the millimeter wave radar platform,this thesis has carried out research on methods such as modeling of millimeter wave radar human head echo signals,millimeter wave radar discrete head movement recognition methods based on multidomain fusion,and millimeter wave radar continuous head movement recognition based on recurrent neural networks.The main work is as follows:1、According to the characteristics of head movements,an experimental scene for detecting human head movements is designed,and data from seven types of head movements are collected.The millimeter wave radar echo signals of human head movements are analyzed and modeled.After processing the echo data,time-range maps,time-frequency representation plots of three kinds of time-frequency analysis methods,and range-doppler maps are obtained.The time-frequency analysis methods include short time fourier transform,wavelet transform,and the smoothed pseudo Wigner-Ville distribution.Establish a data set foundation for subsequent research.2、Aiming at the problem of discrete head movement recognition in millimeter wave radar,a multidomain fusion method is used to study both machine learning and deep learning.The machine learning method combines the feature extraction method of directional gradient histogram;The multidomain fusion deep learning head movement recognition method uses a dual-branch convolutional neural network to simultaneously extract features from data in range domain and time-frequency domain.The feature extraction branch in range domain is a two-dimensional convolutional neural network,and the feature extraction branch in time-frequency domain containing three kinds of time-frequency representation plots is a three-dimensional convolutional neural network that cooperatively extracts three kinds of time-frequency representation plots features.After deeply fusing the network in the medium term,a convolutional block attention module(CBAM)is combined to filter features.Finally,the effectiveness of the algorithm is verified with actual data,and the classification effect is significantly improved compared to machine learning methods.3、Aiming at the problem of millimeter wave radar continuous head movement recognition,a lightweight continuous head movement recognition network based on recurrent neural networks is studied,and the recognition of human continuous head movement is realized.In this network,a deep separable convolution is used to make the model lightweight,and a lightweight channel attention mechanism is combined to enhance the effect of the model.The feasibility and effectiveness of this method for continuous head movement recognition are verified with measured data. |
