1. A multi-target action identification method based on SIMO Doppler radar is characterized by comprising the following steps:

step 1: acquiring gesture action signals of a plurality of targets by using a single-transmitting and multi-receiving Doppler radar, wherein double targets are required to be positioned in the front area of a radar antenna array in the acquisition process;

step 2: carrying out end point detection and signal segmentation on the acquired signals;

and step 3: based on the segmented signal segments, adopting low-pass filtering to filter high-frequency noise, and using discrete wavelet transform to perform data dimension reduction;

and 4, step 4: the method comprises the following steps of separating two different signals by using an independent component analysis algorithm SNRICA based on the maximum signal-to-noise ratio:

1) the radar equipment used by the invention is a single-transmitting double-receiving double-channel Doppler radar, so that the four channels of the preprocessed signals are I1(t), I2(t), Q1(t) and Q2(t), the signals in the real number domain are mapped to the complex number domain, and the order is as follows:

smoothing x (t) to obtain an estimate of x (t)

2) Solving for

3) Solution matrixThen matrix W_estThe optimal estimation of a coefficient matrix W in a blind signal separation problem model y which is Wx is carried out;

4) solving matrix y ═ W_estx, obtaining a separated signal value; where y can be written as:

whereinNamely the separated dual-target motion signals;

and 5: performing two-stage feature extraction on each target respectively, specifically as follows:

firstly, carrying out fast Fourier change on a signal obtained by separation, extracting an envelope curve with the maximum signal intensity from a transformed time-frequency domain graph, then dividing each action into 5 frames by adopting a time window sliding method, extracting characteristic values including a mean value, a variance, a start-stop point slope, a maximum value and a minimum value from a peak envelope of each frame, and calculating the characteristic values to serve as dynamic characteristics of the characteristic values; and restoring the action waveform of each target by adopting a DACM algorithm, wherein the restored action waveform is as follows:

after the action waveform is restored, dividing the image into frames by adopting a sliding window method, and collecting the characteristics including mean value, variance, maximum value, minimum value, positive and negative value accumulation and ratio in each frame; the waveform is then processed to superimpose the phase before each point, i.e.:

similarly, dividing the image into frames by adopting a sliding window method, collecting minimum value, maximum value, start-stop point slope, accumulation sum characteristics in each frame, and combining the characteristics and the waveform extracted by the action waveform into static characteristics;

after the dynamic and static characteristics are extracted, the dynamic characteristics of each action are processed by using an LSTM (least squares TM) to extract deep characteristics on a time domain; then, splicing the processed features and the static features to obtain a feature vector of each target; splicing the feature vectors of the two targets, and then connecting the feature vectors with a full connection layer;

step 6: and taking the extracted feature vectors as training samples, labeling the training samples, inputting the labeled training samples into a classifier for training, acquiring data to be recognized by adopting the same steps, inputting the data into a trained model for processing and classifying.

2. The method of claim 1, wherein the doppler radar in step 1 is RFBeam KLC 7.

3. The method of claim 1, wherein step 2 uses a dual-threshold endpoint detection method to segment the signal segments.

4. The method for identifying multiple target actions based on SIMO Doppler radar as claimed in claim 1, wherein step 3 employs a mean smoothing filter to process the signal: assuming that the new number segment is x (t), the output signal is y (t), and the filtering process uses the formulaAnd (6) processing.

5. The method of claim 1, wherein the classifier is an SVM classifier in step 6.

Background

The document "Lou, Xinye, et al," Gesture-radar: "2018IEEE International Conference on Systems, Man, and Cybernetics (SMC)' IEEE, 2018" proposes a method for completing single target motion recognition based on a single-transmitting and single-receiving Doppler radar, which constructs An interpretable, low-cost and efficient single target gesture recognition scheme based on deep understanding of Doppler radar perception principle, and the document "Pen, Zhenyu, et al," An FMCW radar sensor for human perception in the presence of the single target gesture recognition in the presence of the multi-target targets, 2017First IEEE MTT-S International road interference "MIMO biological Conference (IMBIOC), FM7 finds that the radar range contains radar range information, and performs effective recognition on the scene by using the GPS-based radar, the GPS-based gesture recognition, and the GPS-based recognition Electronics (ICCE), IEEE,2019, "gesture recognition system with interference is realized by innovatively using self-developed 77GHz high-power multiple-input multiple-output (MIMO) FMCW radar," ZHao, Mingmin, et al, "RF-based 3D skeletons," Proceedings of the 2018 Conference of the ACM specific Interest Group on Data communication.2018, "dynamic sensing and modeling of multiple indoor targets are realized for the first time by using an FMCW radar array, and all three documents utilize the high resolution and frequency modulation characteristics of FMCW radar to realize multi-target sensing or other high-precision sensing tasks, but the common Doppler radar based on CW debugging does not have the characteristics. Document "j.ma and x.zhang," bind source segmentation based on maximum signal noise ratio, "in proc.1st int.conf.intel.net.intel.sys" (ICINIS), nov.2008, pp.625-628, "proposes an independent component analysis method based on maximum signal-to-noise ratio, based on which document" Gu, Zhitao, et al, "Remote blank motion segmentation using a single-tone SIMO Doppler radar sensor," IEEE Transactions on geometry Sensing and recovery 57.1 (2018): 462- "innovatively use self-developed single-emission multiple-reception (SIMO) doppler radar to verify the feasibility of the radar in the multi-target identification task, and use the electric slide rail as a test object to successfully separate the periodic motion signals of different slide rails.

The traditional Doppler radar is designed based on the Doppler effect, and the received sensing signal is the Doppler frequency shift of a moving target. The frequency shift can only represent whether an object is close to or far away from the object, but cannot reflect information such as the number, direction, distance and the like of the object. However, the related art does not form a multi-target motion recognition solution for doppler radar, based on the learning and research of these works.

Disclosure of Invention

Technical problem to be solved

In order to avoid the defects of the prior art, the invention provides a multi-target action recognition method based on the SIMO Doppler radar, which can solve the problem that the traditional sensing system based on the Doppler radar does not have the multi-target sensing capability.

Technical scheme

A multi-target action identification method based on SIMO Doppler radar is characterized by comprising the following steps:

step 1: acquiring gesture action signals of a plurality of targets by using a single-transmitting and multi-receiving Doppler radar, wherein double targets are required to be positioned in the front area of a radar antenna array in the acquisition process;

step 2: carrying out end point detection and signal segmentation on the acquired signals;

and step 3: based on the segmented signal segments, adopting low-pass filtering to filter high-frequency noise, and using discrete wavelet transform to perform data dimension reduction;

and 4, step 4: the method comprises the following steps of separating two different signals by using an independent component analysis algorithm SNRICA based on the maximum signal-to-noise ratio:

1) the radar equipment used by the invention is a single-transmitting double-receiving double-channel Doppler radar, so that the four channels of the preprocessed signals are I1(t), I2(t), Q1(t) and Q2(t), the signals in the real number domain are mapped to the complex number domain, and the order is as follows:

smoothing x (t) to obtain an estimate of x (t)

2) Solving for

3) Solution matrix W_est＝[eig(K₂ ^-1K₁)]^*(ii) a Then matrix W_estThe optimal estimation of a coefficient matrix W in a blind signal separation problem model y which is Wx is carried out;

4) solving matrix y ═ W_estx, obtaining a separated signal value; where y can be written as:

whereinNamely the separated dual-target motion signals;

and 5: performing two-stage feature extraction on each target respectively, specifically as follows:

firstly, carrying out fast Fourier change on a signal obtained by separation, extracting an envelope curve with the maximum signal intensity from a transformed time-frequency domain graph, then dividing each action into 5 frames by adopting a time window sliding method, extracting characteristic values including a mean value, a variance, a start-stop point slope, a maximum value and a minimum value from a peak envelope of each frame, and calculating the characteristic values to serve as dynamic characteristics of the characteristic values; and restoring the action waveform of each target by adopting a DACM algorithm, wherein the restored action waveform is as follows:

after the action waveform is restored, dividing the image into frames by adopting a sliding window method, and collecting the characteristics including mean value, variance, maximum value, minimum value, positive and negative value accumulation and ratio in each frame; the waveform is then processed to superimpose the phase before each point, i.e.:

similarly, dividing the image into frames by adopting a sliding window method, collecting minimum value, maximum value, start-stop point slope, accumulation sum characteristics in each frame, and combining the characteristics and the waveform extracted by the action waveform into static characteristics;

after the dynamic and static characteristics are extracted, the dynamic characteristics of each action are processed by using an LSTM (least squares TM) to extract deep characteristics on a time domain; then, splicing the processed features and the static features to obtain a feature vector of each target; splicing the feature vectors of the two targets, and then connecting the feature vectors with a full connection layer;

step 6: and taking the extracted feature vectors as training samples, labeling the training samples, inputting the labeled training samples into a classifier for training, acquiring data to be recognized by adopting the same steps, inputting the data into a trained model for processing and classifying.

Preferably: the Doppler radar in step 1 is RFBeam KLC 7.

Preferably: step 2 uses a double-threshold end-point detection method to segment the signal segment.

Preferably: step 3, processing signals by adopting an average smoothing filter: assuming that the new number segment is x (t), the output signal is y (t), and the filtering process uses the formulaAnd (6) processing.

Preferably: and the classifier in the step 6 is an SVM classifier.

Advantageous effects

The invention provides a multi-target action recognition method based on SIMO Doppler radar, which solves the problem that a perception scheme based on the Doppler radar does not have the capability of multi-target perception in the field of intelligent human-computer interaction and wireless perception. Firstly, acquiring gesture action signals of a plurality of targets by using a single-transmitting and multi-receiving Doppler radar, receiving and recording the acquired signals by an upper computer, and handing the signals to a subsequent algorithm for processing; then, preprocessing the data, including end point detection and signal segmentation, low-pass filtering, discrete wavelet transform and the like; secondly, separating multi-target signals by adopting an independent component analysis method based on the maximum signal-to-noise ratio, then innovatively extracting the feature vectors of all targets by adopting a two-stage feature extraction method, finally training and classifying by using simple classifiers such as SVM and RF, and judging the gesture action of each detected target according to the classification result. In this way, the invention can use doppler radar to identify multi-target motion patterns.

According to the invention, commercial equipment RFBeam Klc7 and ST200 are used for building a sensing platform, and experiments prove that the method can realize multi-target periodic sliding rail motion mode identification and aperiodic gesture motion identification, and the accuracy of multi-target gesture identification of six basic gestures can reach 90%.

Drawings

The drawings are only for purposes of illustrating particular embodiments and are not to be construed as limiting the invention, wherein like reference numerals are used to designate like parts throughout.

FIG. 1 is a frame of a multi-target gesture recognition system based on SIMO Doppler radar in the invention.

FIG. 2 is a diagram of a framework of algorithm sub-modules for two-stage feature extraction in a system framework.

FIG. 3 is an example deployment diagram of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.

The technical scheme adopted by the invention is integrally divided into three parts. Firstly, a single-transmitting and multi-receiving Doppler radar is required to be used for acquiring gesture action signals of a plurality of targets, double targets are required to be located in the front area of a radar antenna array in the acquisition process, the relative positions and angles of the double targets are not strictly required, and the acquired signals are received and recorded by an upper computer and are handed to a subsequent algorithm for processing; in the second part, the invention carries out preprocessing on the data, and comprises methods of endpoint detection, data segmentation, low-pass filtering, discrete wavelet transformation and the like to eliminate noise and carry out data dimension reduction; then, the third part is entered, which is also the core part of the invention, and the multi-target signals are separated by adopting an independent component analysis method based on the maximum signal-to-noise ratio, and then the motion (gesture) state characteristics of each target are respectively extracted by adopting a two-stage characteristic extraction method: firstly, respectively extracting static characteristics and dynamic characteristics of a target, applying a DACM algorithm of motion waveform reduction, a fast Fourier transform FFT and a time-frequency domain analysis method in the extraction process, secondly, performing weighted fusion on the respective characteristics of the multiple targets, and finally training and classifying the processed characteristics by using simple classifiers such as SVM and RF after the characteristics are extracted because the classifiers are not the key points of the invention, and judging the gesture action of each detected target according to the classification result.

Referring to fig. 1, the specific implementation steps of the present invention are as follows:

step one, the initiator of the sensing task deploys the sensing environment according to fig. 3 by a commercial SIMO doppler radar (such as RFBeam KLC7) and a notebook computer. Two persons (perception targets) stand facing the radar at the position about 2m in front of the radar, and when the system normally operates, the two persons respectively perform gesture actions at the same time so as to finish action signal acquisition.

And step two, carrying out end point detection and signal segmentation on the acquired signals. In particular, the present invention uses a dual-threshold endpoint detection method to segment signal segments. The device sampling rate used is 66KHz, the frame size is set to 6000 (i.e. 0.1 seconds long), the frame shift is set to 3000(0.05 seconds), and if each frame is represented as x (i), then the short-time zero-crossing rate SN and the short-time energy SE are respectively:

in the formula, T represents a zero-crossing threshold, which is empirically set to 0.015.

And thirdly, based on the segmented signal segments, adopting low-pass filtering to filter high-frequency noise, and using discrete wavelet transform to perform data dimension reduction. Specifically, the invention uses a mean smoothing filter to process the signal. Assuming that the new number segment is x (t), the output signal is y (t), and the filtering process uses the formulaAnd (6) processing.

For the filtered signal, the discrete wavelet transform is used for data dimension reduction, the iteration number is 8, namely the number of layers of the discrete wavelet transform is 8.

After completing the basic data preprocessing in the fourth step and the first three steps, separating two different signals by using an independent component analysis algorithm SNRICA based on the maximum signal-to-noise ratio in the step, wherein the algorithm steps are as follows:

1) the radar equipment used by the invention is a single-transmitting double-receiving double-channel Doppler radar, so that the four channels of the preprocessed signals are I1(t), I2(t), Q1(t) and Q2(t), the signals in the real number domain are mapped to the complex number domain, and the order is as follows:

smoothing x (t) to obtain an estimate of x (t)

2) Solving for

3) Solution matrix W_est＝[eig(K₂ ^-1K₁)]^*. Then matrix W_estThe optimal estimation of a coefficient matrix W in a blind source signal separation problem model y is Wx;

4) solving matrix y ═ W_estx, to obtain the separated signal value. Where y can be written as:

whereinNamely the separated dual target motion signals.

Step five, filtering the separated signals by using a PCA algorithm to eliminate noise introduced in the signal processing process of the previous step;

and step six, performing two-stage feature extraction on each target respectively by referring to fig. 2.

Firstly, carrying out fast Fourier change on a signal obtained by separation, extracting an envelope curve with the maximum signal intensity from a transformed time-frequency domain graph, then dividing each action into 5 frames by adopting a time window sliding method, extracting characteristic values including a mean value, a variance, a start-stop point slope, a maximum value, a minimum value and the like from a peak envelope of each frame, and calculating the characteristic values to serve as dynamic characteristics of the characteristic values; the DACM algorithm is adopted to restore the action waveform of each target, and the restored action waveform is as follows:

after the action waveform is restored, dividing the image into frames by adopting a sliding window method, and collecting characteristics including mean value, variance, maximum value, minimum value, positive and negative value accumulation, ratio and the like in each frame; the waveform is then processed to superimpose the phase before each point, i.e.:

and similarly, dividing the image into frames by adopting a sliding window method, collecting characteristics including minimum value, maximum value, start-stop point slope, accumulation sum and the like in each frame, and combining the characteristics and the waveform extracted by the action waveform into static characteristics.

After extracting the dynamic and static features, the dynamic features of each action are processed using LSTM to extract the deep features in the time domain. And then, splicing the processed features and the static features to obtain a feature vector of each target. After the feature vectors of the two targets are spliced, the feature vectors are connected with the full-connection layer to achieve the effect of feature information supplement.

And step seven, taking the extracted feature vectors as training samples, inputting the training samples into a simple classifier (such as a random forest) for training after the labels are printed, acquiring data to be recognized by adopting the same steps, and inputting the data into a trained model for processing and classifying.

Therefore, the multi-target action recognition scheme based on the SIMO Doppler radar can be realized.

The invention is a novel technology for multi-target identification based on Doppler radar, makes targeted innovation aiming at the bottleneck of multi-target identification performance of multi-radar, combines research results in the field of blind source signal separation and an adaptive feature extraction scheme, and can realize an efficient multi-target action identification solution only by using a simple classifier.

While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and various equivalent modifications or substitutions can be easily made by those skilled in the art within the technical scope of the present disclosure.