Indoor non-contact human activity recognition method and system
Disclosed are an indoor non-contact human activity recognition method and system. The method comprises: collecting an indoor reflected signal by using an antenna array; filtering the reflected signal to obtain a noise-removed reflection signal; and inputting the noise-removed reflected signal to a pre-trained human activity recognition model, and determining a human activity category, the human activity recognition model being a pre-trained CNN network model based on a transfer learning algorithm. The recognition method and system have the advantages that: the antenna array is configured for collecting human actions to carry out activity recognition indoors, which can be applied to home-based care scenes; original data is denoised, so that most of high-frequency noises can be removed, and a phase change of the signal is reserved; a CNN structure is adopted for training so as to reduce a complexity of the system location-free sensing.
1. An indoor non-contact human activity recognition method, comprising:
collecting an indoor reflected signal by using an antenna array;
filtering the reflected signal to obtain a noise-removed reflection signal; and
inputting the noise-removed reflected signal to a pre-trained human activity recognition model, and determining a human activity category, the human activity recognition model being a pre-trained convolutional neural network (CNN) model based on a transfer learning algorithm;
wherein the reflected signal is filtered by principal component analysis to obtain the noise-removed reflection signal, and the noise-removed reflected signal is inputted to the pre-trained human activity recognition model;
wherein the filtering the reflected signal by principal component analysis comprises:
dividing an indoor activity space into a plurality of position areas, collecting a data packet of a fixed-time reflected signal of each activity in each position area, and removing a static path part in the data packet of the fixed-time reflected signal of each activity;
cutting the data packet of the fixed-time reflected signal of each activity with the static path part removed into a plurality of time slots with a fixed time duration, and forming a matrix H corresponding to the activity according to all the time slots;
obtaining a covariance matrix H T ×H according to the matrix H, and calculating a correlation matrix of the covariance matrix H T ×H, wherein the correlation matrix is composed of correlation coefficients of the covariance matrix, a size of the correlation matrix is N×N, and Nis a number of data streams;
carrying out feature decomposition on the obtained correlation matrix, and calculating to obtain a feature vector; and
carrying out signal reconstruction according to a formula h a =H×q a to obtain the noise-removed reflected signal, wherein q a is an ath feature vector, and h a is an a th principal component.
2. The indoor non-contact human activity recognition method according to claim 1 , wherein a training process of the human activity recognition model comprises:
acquiring a pre-built CNN model, wherein a structure of the CNN model comprises a plurality of convolutional layers, a pooling layer and one fully connected layer, the convolutional layers are sequentially connected through the pooling layer, and the last convolutional layer is connected with the fully connected layer through the pooling layer;
determining a training position according to a conclusion that adjacent positions have similar statistical features and a result of a maximum mean difference, and acquiring the noise-removed reflected signal at the training position to form an initial training data set;
training the CNN model with the initial training data set to obtain initial optimal model parameters;
initializing the CNN model with the initial optimal model parameters;
acquiring the noise-removed reflected signal at a second training position to form a secondary training data set; and
freezing parameters of the convolutional layer and the pooling layer before the fully connected layer, and then training the fully connected layer with the secondary training data set until an upper limit of iteration times or a preset precision is reached, so as to obtain the trained human activity recognition model.
3. An indoor non-contact human activity recognition system, comprising:
a collecting module configured for collecting an indoor reflected signal by using an antenna array;
a filtering module configured for filtering the reflected signal to obtain a noise-removed reflection signal; and
a processing module configured for inputting the noise-removed reflected signal to a pre-trained human activity recognition model, and determining a human activity category, the human activity recognition model being a pre-trained convolutional neural network (CNN) model based on a transfer learning algorithm;
wherein the filtering module is configured for filtering the reflected signal by principal component analysis to obtain the noise-removed reflection signal;
wherein the filtering module comprises:
a division module configured for dividing an indoor activity space into a plurality of position areas, collecting a data packet of a fixed-time reflected signal of each activity in each position area, and removing a static path part in the data packet of the fixed-time reflected signal of each activity;
a cutting module configured for cutting the data packet of the fixed-time reflected signal of each activity with the static path part removed into a plurality of time slots with a fixed time duration, and forming a matrix H corresponding to the activity according to all the time slots;
a first calculating module configured for obtaining a covariance matrix H T ×H according to the matrix H, and calculating a correlation matrix of the covariance matrix H T ×H, wherein the correlation matrix is composed of correlation coefficients of the covariance matrix, a size of the correlation matrix is N×N, and N is a number of data streams;
a second calculating module configured for carrying out feature decomposition on the obtained correlation matrix, and calculating to obtain a feature vector; and
a reconstruction module configured for carrying out signal reconstruction according to a formula h a H×q a to obtain the noise-removed reflected signal, wherein q a is an a th feature vector, and h a is an a th principal component.
4. The indoor non-contact human activity recognition system according to claim 3 , wherein the processing module comprises a model training module configured for:
acquiring a pre-built CNN model, wherein a structure of the CNN model comprises a plurality of convolutional layers, a pooling layer and one fully connected layer, the convolutional layers are sequentially connected through the pooling layer, and the last convolutional layer is connected with the fully connected layer through the pooling layer;
determining a training position according to a conclusion that adjacent positions have similar statistical features and a result of a maximum mean difference, and acquiring the noise-removed reflected signal at the training position to form an initial training data set;
training the CNN model with the initial training data set to obtain initial optimal model parameters;
initializing the CNN model with the initial optimal model parameters;
acquiring the noise-removed reflected signal at a second training position to form a secondary training data set; and
freezing parameters of the convolutional layer and the pooling layer before the fully connected layer, and then training the fully connected layer with the secondary training data set until an upper limit of iteration times or a preset precision is reached, so as to obtain the trained human activity recognition model.