DNN-based human face classification

Abstract

Extracting target information from the radar signal and the measurement environment includes the steps of transmitting a frequency modulated continuous wave transmit signal using a radar, receiving a reflected signal reflected from a human face at multiple antenna elements, multiplying the transmit signal and the reflected signal using a mixer to generate a mixed signal, passing the mixed signal through a low pass filter to generate a baseband signal including a sine wave signal, extracting a frequency of each sine wave signal from the baseband signal to generate extracted frequencies, and measuring the distance between the human face and the radar using the extracted frequencies.

Description

[Technical field]

[0001] The present invention relates to human face classification, and more particularly to deep neural network (DNN) based human face classification. [Background technology]

[0002] Radio detection and ranging (i.e., radar) sensors are widely used in various fields for various purposes. In recent years, the frequency band and bandwidth used by radar sensors have expanded to the millimeter wave band, and there has been an increasing interest in radars with smaller size and higher range resolution. In other words, radar sensors are becoming more miniaturized, and it is expected that radar sensors will be utilized in other applications. Summary of the Invention [Problem to be solved by the invention]

[0003] In this invention, human face classification is described. [Means for solving the problem]

[0004] In one embodiment, a method for extracting target information from a radar signal and a measurement environment is disclosed, the method includes the steps of transmitting a frequency-modulated continuous wave (FMCW) transmit signal using a radar, receiving a reflected signal reflected from a human face at multiple antenna elements, multiplying the transmit signal and the reflected signal using a mixer to generate a mixed signal, passing the mixed signal through a low pass filter to generate a baseband signal including a sinusoidal signal, extracting a frequency of each sinusoidal signal from the baseband signal to generate an extracted frequency, and measuring a distance between the human face and the radar using the extracted frequency.

[0005] In another embodiment, a method is disclosed that includes receiving an input signal from each of a plurality of receive antenna elements of a radar, the input signal being a signal reflected from a face, low pass filtering the input signal to generate a filtered output, passing the filtered output through an analog-to-digital converter and sampling at a sampling frequency to generate a discrete filter output, applying a transformation to the discrete filter output to calculate a distance between the face and the radar, and applying a false alarm rate detection technique to the transformed output to generate a frequency index corresponding to the distance.

[0006] In another embodiment, a non-transitory computer-readable storage medium storing a computer program for extracting target information from a radar signal and a measurement environment is disclosed, the computer program including executable instructions to cause a computer to transmit a frequency modulated continuous wave transmit signal using a radar, receive a reflected signal reflected from a human face at multiple antenna elements, multiply the transmit signal and the reflected signal using a mixer to generate a mixed signal, pass the mixed signal through a low pass filter to generate a baseband signal including a sinusoidal signal, extract a frequency of each sinusoidal signal from the baseband signal to generate an extracted frequency, and measure a distance between the human face and the radar using the extracted frequency.

[0007] Other features and advantages of the present invention will become apparent from the following description which illustrates, by way of example, embodiments of the invention. [Brief description of the drawings]

[0008] The patent or application file contains one or more drawings executed in color. Copies of this patent or patent application publication containing color drawing(s) will be provided by the Office upon request and payment of the necessary fee.

[0009] The details of the present invention, both as to its structure and operation, may be gleaned in part by study of the accompanying drawings, in which like reference numerals refer to like parts.

[0010] [Figure 1] 4 illustrates the time-frequency slope of a transmitted FMCW radar signal in accordance with one embodiment of the present invention. [Figure 2a] 1 shows a millimeter wave FMCW radar according to an embodiment of the present invention. [Figure 2b] 1 illustrates a circuit layout of an FMCW radar including a single transmit antenna and three receive antenna elements according to one embodiment of the present invention. [Figure 3a] Indicates the radar azimuth angle relative to the face. [Figure 3b] Indicates the elevation angle of the radar relative to the face. [Figure 4] 2 shows the structure of an MLP network according to one embodiment of the present invention. [Figure 5a] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5b] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5c] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5d] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5e] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5f]1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5g] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 5h] 1 shows the cumulative F(L3[n]) of eight human faces over a 20 second period when the faces are approximately 30-50 cm away from the center of the radar, according to one embodiment of the present invention. [Figure 6] This shows F(L3[n]) for one subject when the human face is approximately 40 cm away from the radar. [Figure 7] 1 illustrates a data vector for training a DNN according to one embodiment of the present invention. [Figure 8] 1 illustrates a process that utilizes only the signal portion indicated by the red rectangular box according to one embodiment of the present invention. [Figure 9] The composition of FFT indices corresponding to the distances around the face is shown. [Figure 10] 4 shows the average face classification performance for eight different subjects according to one embodiment of the present invention. [Figure 11] 4 shows classification performance according to one embodiment of the present invention. [Figure 12] 4 is a flow chart of a process for extracting target information from a radar signal and a signal measurement environment according to one embodiment of the present invention. [Figure 13] 1 is a flowchart of a process for training a deep neural network (DNN) on a derived radar signal according to one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] As mentioned above, radar sensors are widely used in various fields and for various applications. The frequency bands and bandwidths used by radar sensors have expanded to the millimeter wave band, and there is growing interest in radars with smaller size and higher range resolution. In other words, radar sensors are becoming more miniaturized, and it is expected that radar sensors will be utilized in applications other than existing applications. In one example, radar is integrated into a smartphone for face and gesture recognition. In another example, ultra-wideband (UWB) radar is integrated into a smartphone to find other smartphones.

[0012] Certain embodiments of the present invention provide for classifying human faces using a miniature millimeter wave radar sensor. In one embodiment, the radar sensor transmits a frequency-modulated continuous waveform (e.g., operating in the 61 GHz band) and receives the reflected signal using spatially separated receive antenna elements. Since different people have different face shapes and configurations, the reflection characteristics of the radar signal are also differentiated from each other. Using the signals received from the multiple antenna elements, a deep neural network (DNN) can be trained for human face classification. When training the DNN with real radar signals, many human faces can be classified with high accuracy.

[0013] From such description, it should become apparent how to implement the invention in various embodiments and applications. However, although various embodiments of the invention are described herein, it should be understood that such embodiments are presented by way of example only, and not by way of limitation. Thus, detailed descriptions of various embodiments should not be construed as limiting the scope or breadth of the present disclosure.

[0014] I. Introduction In the present invention, a small radar sensor is proposed to recognize human faces and distinguish other people's faces. Due to the rapid development of machine learning technology, research on target classification using radar sensors has been actively conducted in recent years. For example, various human actions have been classified by applying machine learning technology to radar sensor data. Some have also proposed using automobile radar sensors to identify radar targets such as pedestrians, cyclists, and automobiles. Recently, research is also being conducted to distinguish unmanned aerial vehicles by reflected radar signals. All such target classification research is based on the fact that the reflection attributes of objects are different. Similarly, the shape and structure of human faces are different from person to person, so if an appropriate machine learning technology is applied to radar data, human faces can be sufficiently distinguished. Through face classification using radio waves, the weaknesses of camera-based face recognition methods can be complemented.

[0015] In one embodiment of the present invention, the radar sensor operates at a center frequency of 61 GHz and transmits a frequency modulated continuous wave (FMCW). The signal reflected from a person's face is then received by multiple receive antenna elements. The antenna elements are arranged in elevation and azimuth directions, so that the received signal contains the person's spatial facial features. In addition, each person's face has different components, so the reflection characteristics are also different for each person. Therefore, by concatenating the signals received from the multiple antenna elements and using them as inputs for a deep neural network (DNN), an effective classifier can be designed that can distinguish between different people's faces.

[0016] In one embodiment of the present invention, the classification performance of the proposed DNN-based face classification method is evaluated and compared with other existing feature-based machine learning techniques. The advantage of the proposed DNN-based method is that it does not require a feature extraction stage. The applicability of DNN to radar sensor data has already been confirmed. However, most deep learning-based target classification research has focused on large-scale radar systems that operate in low frequency bands or on targets at relatively long distances.

[0017] The embodiments disclosed in Section II describe signal processing methods for extracting target information from radar signals and the signal measurement environment. The embodiments disclosed in Section III describe methods for training a DNN with the obtained radar signals and the classification results. Section IV includes conclusions.

[0018] II. Target measurement using 61GHz FMCW radar A. Signal processing in FMCW radar In an FMCW radar system, the frequency of the transmitted signal varies linearly with time. Figure 1 shows the time-frequency slope of the transmitted FMCW radar signal, where 1, 2, ..., N are the indices of each frame. Thus, the transmitted signal T(t) of a frame can be expressed as:

number

[0019] Here, A T is the amplitude of the transmitted signal, f c is the center frequency of the modulated signal, ΔB is the operating bandwidth, and ΔT is the sweep time. The transmitted signal is also called an up-chirp signal because the frequency increases rapidly. If the up-chirp signal is reflected from L targets, the received signal R(t) may be expressed as follows:

number

[0020] Here, A Rl (l=1, 2, …, L) is the amplitude of the signal reflected from the first target, and f dl is the Doppler frequency caused by the relative velocity between the first target and the radar, and t dl is the time delay due to the relative range between the first target and the radar. The desired signal containing the information of the first target is d l (t), and n(t) represents the noise added at the receive antenna.

[0021] The transmit signal T(t) is multiplied with the receive signal R(t) by passing it through a frequency mixer. The output of the mixer M(t) is given by:

number

[0022] To extract the baseband signal, M(t) is passed through a low-pass filter. The filter output may be approximated as:

number

[0023] where L(.) represents the low pass filter output.

[0024] Since L(M(t)) is the sum of sinusoidal signals, we apply a Fourier transform to extract the frequency of each signal. The extracted frequency (f^ l ) (where l=1, 2, ..., L) can be expressed as follows:

number

[0025] Here, R l and v l are the relative range and relative velocity between the first target and the radar, respectively, and c is the propagation velocity of the transmitted radar signal.

[0026] When measuring the radar signal reflected from a human face, f^ l is (ΔB / ΔT)t dl (=ΔB / ΔT×2R l / c) that is, since many measurement points on the face are fixed relative to the radar, the target velocity f dl Therefore, the beat frequency can be used to estimate the distance between the face and the radar, which is calculated as follows:

number

[0027] where ΔT, ΔB, and c are already known to the radar system.

[0028] B. Measurement environment Fig. 2a shows a millimeter wave FMCW radar according to an embodiment of the present invention. Fig. 2b shows a circuit layout of an FMCW radar including a single transmitting antenna and three receiving antenna elements according to an embodiment of the present invention.

[0029] As shown in Figures 2a and 2b, FMCW radar utilizes a single transmit antenna and three receive antenna elements (two in azimuth and two in elevation, one in each direction shared by one antenna element). The distance between adjacent antenna elements is 0.5c / f c In the antenna system, Tx is a signal, R x 1. R x 2, and R x 3, the reflected signal is received at the same time. x 1. R x 2, and R x Although only 3) is shown, any number of antenna elements (two or more) may be used in other embodiments. In addition, fc, ΔB, and ΔT in mathematical formula 1 are set to 61 GHz, 6 GHz, and 12.5 ms, respectively. In addition, one frame of the FMCW radar signal has a length of 50 ms, which is composed of a signal transmission time of 12.5 ms and a signal processing time of 37.5 ms.

[0030] Using the radar, an experimental environment is set up as shown in Figures 3a and 3b, and reflected radar signals are collected. Figure 3a shows the azimuth angle of the radar relative to a face, and Figure 3b shows the elevation angle of the radar relative to the face.

[0031] In one exemplary embodiment with eight subjects, the radar signal is measured by changing the distance between the radar and the face from 30 cm to 50 cm. Also, since the reflected signal may vary depending on the azimuth and elevation angles (e.g., θ and Φ in Fig. 3a and Fig. 3b) between the face and the front direction of the radar, the received radar signal is obtained by changing θ and Φ from -5° to +5°.

[0032] III. PROPOSED FACE CLASSIFICATION USING DNN A. Basic principles of DNN The most typical type of DNN structure is the Multilayer Perceptron (MLP), in which each layer is fully connected to the adjacent layers. FIG. 4 shows an exemplary structure of the MLP network. The exemplary structure includes an input layer, multiple hidden layers, and an output layer. Each layer is composed of multiple nodes, and each node is connected to the nodes of the adjacent layers. The network shown in FIG. 4 is trained through an iterative process of forward propagation and backward propagation. In forward propagation, each layer transmits values ​​to the next layer by using weights and activation functions. X i (k) and Y o (k) are called the input and output vectors of the kth layer, and W (k) is called the weight matrix between the kth layer and the (k+1)th layer. In this case, the input vector for the (k+1)th layer can be expressed as follows:

number

[0033] where f a represents the activation function that creates nonlinearity in the network. Meanwhile, the weight values ​​are updated in backpropagation by calculating the slope of the cost function for each weight. Let W be the weight value before backpropagation. b Then, the updated weights W after backpropagation are a is as follows:

number

[0034] Here, α is the learning rate that determines the speed of learning, and J is the cost function that indicates the error between the trained value and the actual value. Both forward and back propagation, called epochs, are performed by using the weight matrix W (k) This is repeated several times to properly train the

[0035] B. Radar signal preprocessing for data learning To train a classifier using DNN, the input format must be determined appropriately. Therefore, the appropriate input is formulated according to the characteristics of the radar signal. The signal received from each receiving antenna element is expressed as R q (t) (q=1, 2, 3), the low-pass filter output for each received signal is L(T(t)R q (t))=L(M q (t)). The low-pass filter output passes through an analog-to-digital converter, and f s (=1 / T s ), the discrete filter output may be expressed as:

number

[0036] Here, N s represents the number of sampled points. As mentioned in Section II-A, the distance between the face and the radar can be extracted by applying a Fast Fourier Transform (FFT) to the sampled signal, F(L q [n]).

[0037] For example, Figure 6 shows F(L3[n]) for one subject when the face is about 40 cm away from the radar. Therefore, the number of FFT points (NF) is N sIt is set equal to. By applying an ordered statistic constant false alarm rate (OS-CFAR) detection algorithm to F(L3[n]), an FFT index corresponding to the face position can be extracted. From Equation 6, in the radar system of the present invention, one FFT index can be calculated to be 2.5 cm. In the above step, extracting the FFT index includes appropriately bounding the output around the distance between the face and the radar (see FIG. 8). It should be noted that if the above outputs are used in series (i.e., concatenated), the network used is classified as a DNN. Alternatively, if the above outputs are used in parallel, it is classified as another deep learning technology called a Convolutional Neural Network (CNN).

[0038] Figures 5a to 5h show F(L3[n]) accumulated for 20 seconds when eight human faces were located approximately 30 to 50 cm away from the center of the radar. The distance between the radar and the human faces can be estimated by converting the FFT index with mathematical formula 6. As shown in Figures 5a to 5h, the strength of the reflected signal is strong at around 40 cm. Figures 5a to 5h show the difference in the received signals for different faces located at the same distance. In addition, there is a phase difference e between the antenna elements located on the azimuth axis. jπsinθ There is a phase difference e between the antenna elements located on the altitude axis. jπsinφ Therefore, F(L1[n]), F(L2[n]), and F(L3[n]) also have different patterns for the same face.

[0039] Generally, the feature vector of the extracted signal can be used as the input of the DNN for the radar data, or the signal in the time domain or frequency domain can be used as a whole. In the present invention, the signals received from all antenna elements are combined into one signal vector and used as the input of the DNN. That is, based on the above observation, the form of the data input for training the DNN is expressed as follows:

number

[0040] As shown in Figure 7, there are three receiving antenna elements, so X i [k] is 1×3N s The data vector is extracted from each up-chirp signal (see Fig. 1). The input data is generated on a frame-by-frame basis as mentioned in Section II-B. For each object, 2,000 input data points (i.e., N=2000) are obtained by changing the measurement conditions (e.g., distance and angle). However, the computational complexity increases when using the entire signal, and the number of input data points is increased by 1. i The vector is not used directly because [k] contains unwanted signals that do not contain target information, so signals near the target are used.

[0041] In the example shown in Figure 8, only the signal portion indicated by the red rectangular box is used. Then, the OS-CFAR detection algorithm is applied to each frame to extract the FFT index corresponding to the face position. The signal around the extracted FFT index is truncated and used as the input to the network, which may be expressed as:

number

[0042] where k tgt is the FFT index corresponding to the distance to the face, and k cutis a distance index that takes into account the distance around the face. In this case, k is set so that the area around the face is covered by about 10 cm. cut is set to 4. Therefore, the refined input X i [k] may be configured as shown in FIG.

[0043] C. Face classification results from DNN As mentioned in Section III-B, 16,000 data points were obtained for 8 subjects. 70%, 15%, and 15% of the total data were used as the training, validation, and test sets, respectively. The structure of the DNN used in the present invention is shown in Figure 4. As mentioned above, X i [k] is used as an input vector for the network, and the output layer of the network is represented as Face 1 to Face 8.

[0044] In general, the factor that has the greatest impact on the classification performance of a DNN structure is the number of hidden layers (N l ), the number of nodes in each layer (N n ), and the type of activation function. Therefore, in order to find a network structure suitable for our radar system, we change the above parameters and compare the classification accuracy. The classification performance is evaluated by first changing the number of hidden layers to 1 (N l = 1) and increasing the number of nodes in the hidden layer.

[0045] Figure 10 shows the average face classification performance for eight different subjects. n It was confirmed that if N is 20 or more, there is no further increase in classification performance. n is fixed at 20, and N l We confirmed the classification performance by changing N l The classification performance was almost converged when N was 5 or more. l and N n The values ​​of are set to 5 and 20, respectively. l and N nWhen the same calculation was performed by changing the order of estimation of N, the values ​​of the two factors were calculated to be almost similar. l and N n The confusion matrix is ​​shown when the mean classification accuracy for the eight subjects is 92% or higher.

[0046] [Table 1]

[0047] Besides DNN, classification performance is evaluated using Support Vector Machine (SVM) and tree-based methods, which are widely used in radar target classification. To train a classifier through SVM, bagging tree, or boosting tree, a data vector (X i In some embodiments, features such as amplitude and Doppler frequency are extracted from the reflected signals to train the classifier. Also, features that indicate statistical properties of the distribution such as mean, variance, skewness, and kurtosis have been used in other applications. The performance of such machine learning techniques varies greatly depending on the features extracted. Thus, in the present invention, the signal of Equation 11 is used as input data for both machine learning techniques to determine the classification performance for the same input data. That is, X i Each sample point in [k] acts as a feature in either of the two classifiers.

[0048] [Table 2]

[0049] Table 2 shows the classification results as follows. The average classification accuracy of these methods was lower than that of the proposed method. In the case of such feature-based machine learning methods, the average classification accuracy can be improved if appropriate features are extracted and used to train the classifier. However, the advantage of the method proposed in the present invention is that it shows high classification performance without extracting such features.

[0050] Also, classification performance was confirmed when signals received by only one antenna element instead of three receive antenna elements were used. That is, the input format may be expressed as:

number

[0051] Where:

number

[0052] FIG 12 is a flow chart of a process 1200 for extracting target information from a radar signal and a signal measurement environment, according to one embodiment of the present invention. As shown in FIG 12, a frequency modulated continuous wave (FMCW) signal is transmitted using a radar at block 1210. At block 1220, the signal reflected from a person's face is received by multiple receive antenna elements. The antenna elements are arranged in elevation and azimuth so that the received signal contains the person's spatial facial features. Also, each person has different facial components, resulting in different reflection characteristics.

[0053] In block 1230, the transmit signal is multiplied with the receive signal using a mixer to generate a mixed signal. Then, in block 1240, the mixed signal is passed through a low pass filter to generate a baseband signal including the sinusoidal signal. In block 1250, the frequency of each sinusoidal signal is extracted from the baseband signal. In one embodiment, the frequency of each sinusoidal signal is extracted by applying a Fourier transform. Then, in block 1260, the distance between the face and the radar is measured using the extracted frequency. The signals received from the multiple antenna elements are concatenated, and then the concatenated signal is used as an input to a deep neural network (DNN) to distinguish and / or recognize each human face.

[0054] In one embodiment, a method for extracting target information from a radar signal and a measurement environment is disclosed, the method includes transmitting a frequency modulated continuous wave transmit signal using a radar, receiving a reflected signal reflected from a human face at multiple antenna elements, multiplying the transmit signal and the reflected signal using a mixer to generate a mixed signal, passing the mixed signal through a low pass filter to generate a baseband signal including a sinusoidal signal, extracting a frequency of each sinusoidal signal from the baseband signal to generate extracted frequencies, and measuring a distance between the human face and the radar using the extracted frequencies.

[0055] In one embodiment, the method includes receiving signals from multiple antenna elements and combining the signals. In one embodiment, the method further includes inputting the combined signals into a deep neural network for distinguishing and recognizing human faces. In one embodiment, the deep neural network includes a multi-layer perceptron. In one embodiment, the multiple antenna elements are arranged in elevation and azimuth such that the reflected signals include spatial facial features of human faces. In one embodiment, extracting the frequency of each sinusoidal signal includes applying a Fourier transform. In one embodiment, the transmit signal is an up-chirp signal that rapidly increases in frequency. In one embodiment, the mixer is a frequency mixer. In one embodiment, the radar is a millimeter wave frequency modulated continuous wave radar. In one embodiment, extracting the frequency of each sinusoidal signal includes appropriately bounding each sinusoidal signal around a distance.

[0056] In another embodiment, a non-transitory computer-readable storage medium storing a computer program for extracting target information from a radar signal and a measurement environment is disclosed, the computer program including executable instructions to cause a computer to transmit a frequency modulated continuous wave transmit signal using the radar, receive a reflected signal reflected from a human face at multiple antenna elements, multiply the transmit signal and the reflected signal using a mixer to generate a mixed signal, pass the mixed signal through a low pass filter to generate a baseband signal including a sinusoidal signal, extract a frequency of each sinusoidal signal from the baseband signal to generate an extracted frequency, and measure a distance between the human face and the radar using the extracted frequency.

[0057] In one embodiment, the storage medium further comprises executable instructions for causing the computer to receive signals from multiple antenna elements and combine the signals. In one embodiment, the storage medium further comprises executable instructions for causing the computer to input the combined signals into a deep neural network for distinguishing and recognizing human faces. In one embodiment, the deep neural network comprises a multi-layer perceptron. In one embodiment, the executable instructions for causing the computer to extract the frequency of each sinusoidal signal comprise executable instructions for causing the computer to apply a Fourier transform.

[0058] FIG. 13 is a flow chart of a process 1300 for training a deep neural network (DNN) with a derived radar signal according to one embodiment of the present invention. As shown in FIG. 13, in block 1310, an input signal is received from each receive antenna element. In block 1320, each received signal is low-pass filtered to generate a filtered output. Then, in block 1330, the filtered output is passed through an analog-to-digital converter (ADC) and sampled at a sampling frequency to generate a discrete filter output. In block 1340, a transform is applied to the discrete filter output to generate a transformed output and calculate the distance between the face and the radar. Then, in block 1350, a false alarm rate detection technique is applied to the transformed output to generate a frequency index corresponding to the location of the face. In one embodiment, the transform is a fast Fourier transform (FFT), and the frequency index is an FFT index.

[0059] In one embodiment, a method is disclosed that includes receiving an input signal from each of a plurality of receive antenna elements of a radar, the input signal being a reflected signal from a face, low pass filtering the input signal to generate a filtered output, passing the filtered output through an analog-to-digital converter and sampling at a sampling frequency to generate a discrete filter output, applying a transformation to the discrete filter output to calculate a distance between the face and the radar, and applying a false alarm rate detection technique to the transformed output to generate a frequency index corresponding to the distance.

[0060] In one embodiment, the method further comprises inputting the frequency index into a deep neural network for training. In one embodiment, the deep neural network comprises a multi-layer perceptron including multiple layers, each layer being fully connected to adjacent layers. In one embodiment, the transform is a Fast Fourier Transform. In one embodiment, the frequency index is a Fast Fourier Transform index.

[0061] IV. Conclusion In the present invention, multiple subject faces are classified using data obtained from a small 61 GHz FMCW radar sensor. First, a radar signal is obtained by changing the distance and angle between the radar and the subject's face. The combined signals received from spatially independent receiving antenna elements in one frame are placed together as the input of the DNN. The subject's face was classified with an accuracy of over 92% using the method proposed in the present invention. The performance of the proposed method was then compared with feature-based machine learning methods such as SVM or tree-based methods. The proposed method showed better classification accuracy. It was also confirmed that the face classification performance was better when multiple antenna elements were used. The weaknesses of camera-based face recognition methods were significantly compensated for through the proposed face classification using a small radar sensor.

[0062] The description of the disclosed embodiments is provided to enable one skilled in the art to make or use the present invention as described in the specification set forth above. Various modifications to such embodiments will be apparent to those skilled in the art, and the general principles described herein may be applied to other embodiments without departing from the spirit or scope of the present disclosure. Thus, the technology is not limited to the specific examples described above. Therefore, the description and drawings presented herein are intended to illustrate presently possible embodiments of the present invention, and therefore to illustrate the broadly contemplated claimed subject matter of the present invention. Moreover, the scope of the present disclosure fully encompasses other embodiments that may become apparent to those skilled in the art, and therefore, the scope of the present disclosure is not intended to be limited by anything other than the scope of the appended claims.

Claims

1. 1. A method for extracting target information from a radar signal, the method comprising: training a deep neural network using signals for a plurality of human faces received from the antenna elements; transmitting a frequency-modulated continuous wave transmission signal using the radar; receiving a reflected signal at multiple antenna elements reflected from a human face; combining the received reflected signals; multiplying the transmitted signal and the reflected signal using a mixer to generate a mixed signal; passing the mixed signal through a low pass filter to generate a baseband signal comprising a sinusoidal signal; extracting the frequency of each sinusoidal signal from the baseband signal; measuring a distance between the human face and the radar using the extracted frequency; truncating the combined reflected signals based on the measured distance; inputting the truncated reflection signals into the deep neural network to distinguish a particular face among the plurality of faces; A method for extracting target information, comprising:

2. 2. The method of claim 1, wherein the deep neural network includes a multi-layer perceptron.

3. The method of claim 1 , wherein the multiple antenna elements are arranged in elevation and azimuth such that the reflected signal includes spatial facial features of the person's face.

4. 2. The method of claim 1, wherein extracting a frequency of each of the sinusoidal signals comprises applying a Fourier transform.

5. 2. The method for extracting target information as claimed in claim 1, wherein the transmission signal is an up-chirp signal that rapidly increases in frequency.

6. The method for extracting target information according to claim 1 , wherein the mixer is a frequency mixer.

7. 2. The method for extracting target information as claimed in claim 1, wherein said radar is a millimeter-wave frequency modulated continuous wave radar.

8. A method for detecting a plurality of human faces comprising: training a deep neural network using signals for a plurality of human faces received from a receiving antenna element; receiving an input signal from each of a plurality of receiving antenna elements of a radar, the input signal being a signal reflected from a face; concatenating the received input signals; low-pass filtering the combined input signal to generate a filtered output; passing the filtered output through an analog-to-digital converter and sampling at a sampling frequency to generate a discrete filtered output; applying a transform to the discrete filter output to calculate a distance between the face and the radar; truncating the concatenated input signals based on the calculated distance; inputting the truncated input signal into the deep neural network to distinguish a particular face among the plurality of faces; and applying a false alarm rate detection technique to the transformed output to generate a frequency index corresponding to the distance.

9. The method of claim 8 , further comprising inputting the frequency index into the deep neural network for training.

10. 10. The method of claim 8, wherein the deep neural network comprises a multi-layer perceptron including multiple layers, each layer being connected to adjacent layers.

11. The method of claim 8 , wherein the transform is a Fast Fourier Transform.

12. The method of claim 8 , wherein the frequency index is a Fast Fourier Transform index.

13. 1. A non-transitory computer-readable storage medium storing a computer program for extracting target information from radar signals, the computer program being configured to: training a deep neural network using signals for a plurality of human faces received from the antenna elements; Using the radar to transmit a frequency modulated continuous wave transmission signal; receiving a reflected signal reflected from a human face at multiple antenna elements; Concatenating the received reflected signals; multiplying the transmitted signal and the reflected signal utilizing a mixer to generate a mixed signal; passing the mixed signal through a low pass filter to generate a baseband signal comprising a sinusoidal signal; Extracting the frequency of each sinusoidal signal from the baseband signal; measuring a distance between the human face and the radar using the extracted frequency; Truncate the combined reflected signal based on the measured distance; inputting the truncated reflection signals into the deep neural network to distinguish a particular face among the plurality of faces; A non-transitory computer-readable storage medium containing executable instructions.

14. 14. The non-transitory computer-readable storage medium of claim 13, wherein the deep neural network comprises a multi-layer perceptron.

15. 14. The non-transitory computer-readable storage medium of claim 13, wherein the executable instructions that cause the computer to extract a frequency of each of the sinusoidal signals include executable instructions that cause the computer to apply a Fourier transform.

Images