Through-the-Wall Micro-Doppler De-Wiring Technique via Cycle-Consistent Adversarial Network

Abstract

The radar penetrating technique has aroused a keen interest in the research community, due to its superior abilities for through-the-wall indoor human motion monitoring. Micro-Doppler signatures in this situation play a significant role in recognition and classification for human activities. However, the live wire buried in the wall introduces additive clutters to the spectrograms. Such degraded spectrograms drastically affect the performance of behind-the-wall human activity detection. In this paper, an ultra-wideband (UWB) radar system is utilized in the through-the-wall scenario to get the feature enhanced micro-Doppler signature called range-max time-frequency representation (R-max TFR). Then, a recently introduced Cycle-Consistent Generative Adversarial Network (Cycle GAN) is employed to realize the end-to-end de-wiring task. Cycle GAN can learn the mapping between spectrograms with and without the live wire effect. To minimize the wiring clutters, a loss function called identity loss is introduced in this work. Finally, the proposed de-wiring approach is evaluated through classification. The results show that the proposed Cycle GAN architecture outperforms other state-of-art de-wiring methods.

1. Introduction

Recently, radar based indoor human activity detection has attracted considerable attention from researchers and practitioners due to its outstanding penetration capabilities and robustness against different deployment and illumination conditions [1]. The rising interest is driven by the superb machine learning technique for detection, and the fact that the increasing aging population requires complete elderly care with privacy-preserving indoor monitoring technologies [2,3]. Micro-Doppler signature has been the mainstream feature for recognition and classification since it was introduced by Dr. V. C. Chen to depict micro-motion in the radar observation [4,5]. It thus has been extensively utilized and successfully applied in analyzing the individual motions in free space. Besides, for another rather significant application scene, through-the-wall (TW) indoor human activity monitoring, many unsolved problems still attract researchers’ attentions. The feasibility of measuring reliable TW fall signatures is demonstrated in [6]. The authors in [7] reported a finite-difference time-domain (FDTD) method based simulation approach to generate the micro-Doppler signatures of human moving behind walls. In [8], the influence of the wall, such as the blockage, attenuation, and dispersion of the electromagnetic (EM) waves, was analyzed for the human motion detection task. An ultra-wideband (UWB) bio radar working at L and S bands was developed in [9] to realize penetrating indoor human motion detection. The system has such advantages as strong anti-jamming capability, excellent penetrability, and high-range resolution. A high-resolution range profile was first obtained by using the UWB radar system. Then, a finite impulse response (FIR) high-pass filter was employed to suppress the effect of clutter and wall. However, due to the existence of wall, the micro-Doppler signature become much fainter when compared with those in the free space. Moreover, the wall often introduces extra noises and multipath effects. To remove these effects and obtain the feature enhanced micro-Doppler signatures, a method called the R-max TFR was proposed in [10].

In many realistic TW senarios, live wires are commonly buried in the wall. In China, the live wire typically carries 50 Hz alternative current. The existence of the live wire would bring in additive new micro-Doppler signatures in the spectrogram at fixed frequencies [11]. Such degradation of the spectrograms leads to the failure of a detector, since deep learning methods are sensitive to the abnormal inputs [12].

An intuitive strategy of mitigating the live wire effect is to use clutter reduction techniques [11]. The constant false alarm rate (CFAR) method was adopted to clean the spectrograms in [13,14]. This ideal adaptive thresholding strategy assumes a constant false alarm rate. The Gamma-correction (Gc) method comes with low computational complexity [15], in which the intensity of each pixel I is modified to a new $\widehat{I}=I^{\gamma }$ using a factor $\gamma$. However, Gc method has the drawbacks of attenuating the critical micro-Doppler components in the spectrogram. Both methods thus affect the performance of the subsequent motion recognition [16,17]. In [11], a conditional generative adversarial network (cGAN) [18] was applied to learn the mapping from the live wire corrupted spectrograms to the wiring effect free ones. However, cGAN needs paired training samples, which means that complicated modeling is required to generate the synthetic paired spectrograms.

To overcome the drawback of cGAN, this work turns to an unsupervised learning strategy using a new generative model called Cycle-Consistent Generative Adversarial Network (Cycle GAN). The network can learn the translation between the live wire corrupted and wire clutter free spectrograms in an unsupervised manner without using paired training samples [19]. It has been utilized in the radar image processing scope such as the translation between the optical images and the synthetic aperture radar (SAR) images [20]. The effect analysis of adjusting the key parameters and the size of the dataset for the Cycle GAN are exhibited for better image translation results [21]. The cycle GAN consists of four models: two generators $(G_f,G_b)$ and two discriminators $(D_X,D_Y)$. Define the live wire corrupted spectrogram domain as Y, and the wire clutter free spectrogram as X, then $G_f$ acts as a mapping function $G_f:Y\to X$ to translate the input live wire corrupted spectrogram to the de-wired one. $D_X$ then tries to distinguish whether the spectrogram is from X or $G_f\left(Y\right)$. Because mapping is highly under-constrained, we couple it with the inverse mapping function $G_b:X\to Y$. Through the inverse mapping function, the cycle consistency is formed and serves as a way to link the correspondence between X and Y. At last, a cycle consistency loss function from the orignal Cycle GAN [19] is imposed to realize $G_b\left(G_f\left(Y\right)\right)\approx Y$ and vice versa. To preserve the most authentic human motion micro-Doppler signatures between X and Y, we introduce an additional identity mapping loss function [22], which was first employed in Domain Transfer Network (DTN) to encourage the mapping to keep the color composition of the output as the same as that of the input. The real live wire corrupted and wire clutter free datasets are used for training and testing. For comparison, we also create a dataset of synthetic live wire corrupted spectrograms [11] for training the de-wiring cGAN. Finally, the classification comparisons between the proposed de-wiring method and other state-of-art de-wiring methods are presented in this work.

The contributions of this paper are listed as follows:

(a) We utilize a Cycle GAN-based optimization framework to meet the challenging micro-Doppler signature de-wiring problem without using the complicated wire effect modeling methods.

(b) The identity mapping loss is introduced to ensure more human motion micro-Doppler components are being preserved.

The rest of the paper is organized as follows. A brief introduction to the TW micro-Doppler signature and the live wire effect are presented in Section 2. In Section 3, the principle, architecture, and loss function of the Cycle GAN based de-wiring technique are described. In Section 4, the experimental setup and dataset preparation are introduced. The classification results for the de-wired spectrograms are discussed in comparison with other de-wiring methods. Section 5 summarizes the concluding remarks.

2. Live Wire Effect of TW Indoor Human Activities

2.1. Enhanced Micro-Doppler Signature of TW Indoor Human Activities

We adopt the stepped frequency continuous wave (SFCW) waveform in our UWB system. After the range compression process, the 2D range map (RM) is obtained, whose column and row correspond to the range and slow time, respectively. We utilize the high-pass FIR filter to get rid of the effect of the wall [10]. To get the micro-Doppler signature of human motions, we apply the short time Fourier transformation (STFT) to each range bin of RM over a period of slow-time, and then we can obtain the radar data cube (RDC) [23], which is defined as:

$$\mathrm{RDC}(m,n,k)={\left|\sum _{l=0}^{L-1}\mathrm{RM}(m,n+l)h\left(l\right)e^{-j2\pi kl/L}\right|}^2,$$

(1)

where m, n, and k indicate the range, slow time, and frequency domain correspondingly, l and L represents the index and the total number of the sliding window of STFT, respectively. and $h\left(l\right)$ is a window function. In this paper, all the RDCs are generated by using a Hanning window of length 32, with an overlap of 31 samples. An example of RDC is depicted in Figure 1a for falling.

To efficiently exploit the range information, we search in each range bin of RDC with the fixed slow-time along the frequencies for the maximum values and then project them onto the time-frequency domain, which is called R-max TFR [10], and can be expressed as,

$$\mathrm{R}\text{-}max\mathrm{TFR}(k,n)=\underset{m}{max}\left[\mathrm{RDC}\right(m,k,n\left)\right].$$

(2)

Figure 1b shows an example of R-max TFR for falling. Since the most informative signature is preserved, R-max TFR achieves the highest performance in detection and classification among all the TW micro-Doppler signatures [10].

2.2. Micro-Doppler Signature of Live Wire

In spite of the satisfying classification results of the R-max TFR, once there is a live wire buried in the wall, the corresponding spectrogram would deteriorate sharply [11]. As depicted in Figure 2, the living cable will introduce an extra component of the micro-Doppler signature to the spectrogram.

In order to get rid of the micro-Doppler signature of live wire in the spectrogram, we previously used cGAN technique, which requires paired spectrograms for training. Additionally, we modeled the live wire effect as the impulse response of the sliding window of STFT along with a certain frequency [11]. However, the synthetic spectrogram is not perfectly aligned with the spectrogram with real live wire, which may hinder the performance of de-wiring technique for the real-world one.

3. Cycle GAN Methodology for De-Wiring

To better leverage the live wire free spectrogram, we utilize the Cycle GAN method to transfer micro-Doppler signatures of the wire corrupted spectrograms to those of the clean ones. In this section, the architecture and loss function of the de-wiring Cycle GAN method are presented.

3.1. Formulation of De-Wiring Cycle Generative Adversarial Network

The goal of the de-wiring Cycle GAN is to learn the mapping functions between the live wire corrupted spectrogram dataset Y and the live wire free spectrogram dataset X, given training samples ${\left\{x\right\}}_{i=1}^N$, $x_i\in X$ and ${\left\{y\right\}}_{i=1}^N$, $y_i\in Y$. The spectrogram is denoted as $x\sim p_{spec}\left(x\right)$ and $y\sim p_{spec}\left(y\right)$, where $p_{spec}\left(x\right)$ and $p_{spec}\left(y\right)$ stand for the probability distributions of x and y, correspondingly. As Figure 3 shows, the two mapping functions, forward generator $G_f:Y\to X$ and backward generator $G_b:X\to Y$ are included in the de-wiring Cycle GAN. Furthermore, the two adversarial discriminators are introduced: The first discriminator $D_X$ aims to distinguish between the clean spectrogram $\left\{x\right\}$ and the de-wired spectrogram $\left\{G_f\left(y\right)\right\}$. Likewise, the second discriminator $D_Y$ aims to distinguish between the real live wire spectrogram $\left\{y\right\}$ and the synthetic spectrogram $\left\{G_b\left(x\right)\right\}$. Different from the classical Cycle GAN whose objective only includes the adversarial loss [24] and cycle consistency loss [19], the proposed de-wiring Cycle GAN method further adopts identity loss [22] to encourage the mapping functions to preserve the useful micro-Doppler signatures of human motions.

Basic Cycle GAN contains two kinds of terms: the adversarial loss [18] to update the parameters of the generator and discriminator adversarially; and a cycle consistency loss [19] for preventing the learned mapping functions $G_f$ and $G_b$ from contradicting each other. We further employ the identity loss [22] to maintain the human micro-Doppler signatures in the de-wired spectrograms. Next, the details of the above three kinds of loss functions are presented.

(a) Adversarial loss: The mapping function $G_f:Y\to X$ can be largely improved by tuning it with the corresponding discriminator. We express this objective as the adversarial loss:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{adv}(G_f,D_X,X,Y)& ={\mathbb{E}}_{x\sim p_{spec}\left(x\right)}[log\left(D_X\left(x\right)\right)]\hfill \\ & +{\mathbb{E}}_{y\sim p_{spec}\left(y\right)}[1-log\left(D_Y\left(y\right)\right)],\hfill \end{array}$$

(3)

in which $G_f$ aims to minimize the adversarial loss, while $D_X$ aims to maximize it. $\mathbb{E}[·]$ means the mathematical expectation of the inner function. The similar adversarial loss for backward mapping function $G_b:X\to Y$ can be expressed in the same way: ${\mathcal{L}}_{adv}(G_b,D_Y,X,Y)$.

(b) Cycle consistency loss: To further implement a kind of “meta-supervision” to the mapping functions, Cycle GAN adopts the cycle-consistent constraint [19]. As Figure 3 illustrates, the spectrogram with live wire should be brought back to the original wire-corrupted one by the forward spectrogram translation cycle, i.e., $y\to G_f\left(y\right)\to y_{rec}\approx y$, where $y_{rec}=G_b\left(G_f\left(y\right)\right)$ denotes the rectified spectrogram of y. Equivalently, the same cycle consistency should be satisfied by the backward spectrogram translation cycle: $x\to G_b\left(x\right)\to x_{rec}\approx x$, where $x_{rec}=G_f\left(G_b\left(x\right)\right)$ denotes the rectified spectrogram of x. This behavior can be facilitated by utilizing the cycle consistency loss:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{cycle}(G_f,G_b)& ={\mathbb{E}}_{y\sim p_{spec}\left(y\right)}[\parallel G_b\left(G_f\left(y\right)\right)\parallel ]\hfill \\ & +{\mathbb{E}}_{x\sim p_{spec}\left(x\right)}[\parallel G_f\left(G_b\left(x\right)\right)\parallel ].\hfill \end{array}$$

(4)

(c) Identity loss: Solely considering the original loss functions may result in deterioration of the desired micro-Doppler signatures. We thus employ the identity loss [22] to ensure the maintaining of the micro-Doppler signature in the de-wired spectrogram. This loss can be depicted as:

$$\begin{array}{cc}\hfill {\mathcal{L}}_{ide}(G_f,G_b)& ={\mathbb{E}}_{y\sim p_{spec}\left(y\right)}\left[{\parallel G_f\left(y\right)-y\parallel }_1\right]\hfill \\ & +{\mathbb{E}}_{x\sim p_{spec}\left(x\right)}[\left[{\parallel G_b\left(x\right)-x\parallel }_1\right].\hfill \end{array}$$

(5)

Now we combine the adversarial loss, cycle consistency loss, and identity loss into the final de-wiring Cycle GAN objective:

$$\begin{array}{c}\hfill \mathcal{L}(G_f,G_b,D_Y,D_X)={\mathcal{L}}_{adv}(G_f,D_X,X,Y)\\ \hfill +{\mathcal{L}}_{adv}(G_b,D_Y,X,Y)+\lambda {\mathcal{L}}_{cycle}+\mu {\mathcal{L}}_{ide}(G_f,G_b),\end{array}$$

(6)

where $\lambda$ and $\mu$ denote the weights for ${\mathcal{L}}_{cycle}$ loss and ${\mathcal{L}}_{ide}$ loss, respectively. The mapping function $G_f$ can be obtained by,

$${G_f}^{*}=arg\underset{G_f,G_b}{min}\underset{D_Y,D_X}{max}\mathcal{L}(G_f,G_b,D_Y,D_X).$$

(7)

where ${G_f}^{*}$ denotes one typical solution of $G_f$.

In Section 4, we compare the de-wiring results of the proposed combined loss mapping function with the original one without the identity loss. Results show that the combined loss mapping function is sufficient to regularize the training for a better de-wired spectrogram.

3.2. Network Architecture for Spectrogram De-Wiring Cycle Generative Adversarial Network

Aligned with the purpose of the de-wiring spectrogram task, the forward generator of Cycle GAN should own the ability to restore the micro-Doppler details of human motion while removing the live wire effect. Thus, the network architecture should be appropriately modeled for generating a satisfying de-wired spectrogram.

We adopt the same architecture as [19] for the de-wiring generative networks. As Figure 4 shows, the generator contains three convolution layers with stride-2, nine residual blocks [22], and two fractionally strided convolution layers (transposed convolution blocks) with half stride. Each residual block consists of two reflection padding layers, which are used to reduce the artifacts. Besides, it contains two convolution layers, two instance normalization layers, and a ReLU activation. The forward generator $G_f$ tries to map the single channel input spectrogram with live wire to the output spectrogram without the live wire effect. Similarly, the backward generator $G_b$ tries to map the clean spectrogram to the live wire corrupted spectrogram. After training, we only take the $G_f$ as the de-wiring mapping function.

To stimulate the generator to get better results, we adopt PatchGANs [19] as the discriminators to be trained adversarially against the generators. As Figure 4 illustrates, PatchGAN is in a fully convolutional fashion. Through this, the overlapping spectrogram patches are classified as real or fake. The final judgment output of PatchGAN is determined by averaging all the classification results of patches across the whole spectrogram.

4. Experimental Setup and Results

4.1. Experimental Setup

Through-the-wall radar human motion measurements were conducted to demonstrate the contribution of the proposed Cycle GAN de-wiring technology, which requires both the dataset of the spectrogram with live wire, and the dataset of live wire free spectrogram. We adopt the same UWB radar platform in [9], where two transmitting and four receiving planar logarithmic spiral antennas are utilized to generate and receive a UWB SFCW signal to sense the detection region of interest. The working frequencies of the UWB radar system used in the experiment are from 0.4 GHz to 4.4 GHz, with a step of 5 MHz. The pulse repetition frequency (PRF) is 113 Hz, with the sensitivity of −90 dBm. Firstly, to obtain the dataset X, we conduct experiments of different human motions in through-the-wall environment as shown in Figure 5a. The clean spectrogram dataset X contains nine human motions of two subjects: walking forward (motion 1, M1) and walking backward (motion 2, M2), sitting down (motion 3, M3), standing up from seated (motion 4, M4), fetching down and up (motion 5, M5), crawling forward (motion 6, M6), crawling backward (motion 7, M7), falling forward (motion 8, M8) and falling backward (motion 9, M9). The intuitive illustration of these motions can be found in [10]. For each motion, we have a total number of 1000 samples. Secondly, to attain the dataset Y, we conduct experiments of the same nine human motions in through-the-wall scenario with live wire as illustrated in Figure 5b, and each category contains 1000 entries. Specifically, we turned the lights from off to on to introduce the live wire effect. Some spectrograms of different actions with real-world live wire (input spectrograms) are shown in Figure 6a,g,m,s. In the training stage of the de-wiring Cycle GAN, 80% of both the recorded clean spectrogram dataset X and the live wire spectrogram dataset Y are utilized, whereas the remaining 20% of Y are used to test the effect of the de-wiring Cycle GAN. Besides, the dataset X is also used to generate the the paired wire-corrupted and wire-free spectrograms for training the de-wiring cGAN for comparison.

4.2. De-Wiring Results of Cycle GAN

We perform qualitative and quantitative comparisons of the proposed de-wiring Cycle GAN with many state-of-the-art spectrogram de-wiring methods, such as 2D CA-CFAR, Gamma-correction (Gc) and de-wiring cGAN [11]. Additionally, to state the importance of identity loss, we run the original Cycle GAN to isolate the effect of the identity term.

To visually demonstrate the performance obtained by the proposed de-wiring Cycle GAN, results of four kinds of human motion spectrograms (fetching down and up, sitting down, standing up and falling backward) are presented in Figure 6. We can see that both of the Gc and CA-CFAR fail to remove the live wire component thoroughly, meanwhile changing the desired human motion micro-Doppler signatures. The de-wiring cGAN can reduce most of the live wire micro-Doppler streaks. However, the details of the human motion micro-Doppler signatures may be faded or removed, because of the imprecise modeling method [11]. In comparison, it can be observed that both Cycle GAN methods can remove most of the live wire micro-Doppler effects while maintaining the details of the human motion micro-Doppler signatures. Moreover, thanks to the feature maintaining charater of identity loss, with which the Cycle GAN architecture can suppress the live wire effects thoroughly.

4.3. Classification Results

In this part, the classification results of the aforementioned de-wiring methods are presented, to verify that the novel Cycle GAN with identity loss could achieve the best performance in terms of the classification accuracy of the nine types of the live wire corrupted human motions.

VGG16 [25] is employed to discriminate different de-wired spectrograms of behind wall motions, which is widely used as an effective architecture in the image classification task. The spectrograms of nine human motions (9000) without the live wire effects are employed to train the classifier. As mentioned in Section 4. A, 80% of the live wire spectrogram dataset Y is used for training the Cycle GAN, while the remaining 20% is de-wired by the aformentioned methods and then utilized as the test set of the classifier.

The confusion matrices for test datasets by different de-wiring methods are given in

Table 1. Confusion matrix for none.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	0	0	100%	0	0	0	0	0
M5	0	0	0	0	96%	0	0	4%	0
M6	0	48%	0	0	0	52%	0	0	0
M7	0	0	0	0	0	0	100%	0	0
M8	7%	0	0	0	0	0	0	93%	0
M9	0	0	0	0	0	0	0	0	100%

,

Table 2. Confusion matrix for Gc.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	0	0	100%	0	0	0	0	0
M5	0	0	0	0	81%	0	0	19%	0
M6	0	0	0	0	0	100%	0	0	0
M7	0	0	0	0	0	0	100%	0	0
M8	1%	6%	0	0	0	0	0	94%	0
M9	0	0	0	0	0	0	0	0	100%

,

Table 3. Confusion matrix for CA-CFAR.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	0	0	100%	0	0	0	0	0
M5	0	0	0	0	96%	0	0	0	4%
M6	0	48%	0	0	0	52%	0	0	0
M7	0	77%	0	0	0	0	23%	0	0
M8	3%	0	0	0	1%	0	0	92%	4%
M9	0	0	0	0	0	0	0	0	100%

,

Table 4. Confusion matrix for cGAN.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	0	0	100%	0	0	0	0	0
M5	0	0	0	0	100%	0	0	0%	0
M6	0	100%	0	0	0	0	0	0	0
M7	0	0	0	0	0	0	100%	0	0
M8	6%	0	0	0	0	0	0	94%	0
M9	0	0	0	0	0	0	0	0	100%

,

Table 5. Confusion matrix for the Orignal Cycle GAN.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	3%	0	97%	0	0	0	0	0
M5	0	0	0	0	28%	0	0	72%	0
M6	49%	0	0	0	0	51%	0	0	0
M7	0	0	0	0	0	0	100%	0	0
M8	6%	0	0	0	0	0	0	94%	0
M9	0	0	0	0	0	0	0	0	100%

and

Table 6. Confusion matrix for the proposed method.

	M1	M2	M3	M4	M5	M6	M7	M8	M9
M1	100%	0	0	0	0	0	0	0	0
M2	0	100%	0	0	0	0	0	0	0
M3	0	0	100%	0	0	0	0	0	0
M4	0	0	0	100%	0	0	0	0	0
M5	0	0	0	0	100%	0	0	0	0
M6	0	0	0	0	0	100%	0	0	0
M7	0	0	0	0	0	0	100%	0	0
M8	6%	0	0	0	0	0	0	94%	0
M9	0	0	0	0	0	0	0	0	100%

. The rows and columns of the confusion matrix represent the actual and predicted motion classes, respectively. As presented in Table 1, Table 3, Table 4 and Table 5, sometimes the classifier recognizes the crawling forward to the walking backward or forward in error, because the signal intensity and micro-Doppler signatures of these motions are similiar. Note that the CA-CFAR method even makes negative contribution to the prediction of the VGG16, as is shown in Table 3, possibly because of its attenuation to the micro-Doppler signature. Table 6 has a better classification accuracy than that of Table 5. The result emphasizes the necessity of the identity loss in the de-wiring task. The classification result of the de-wired spectrograms by Cycle GAN with identity loss reaches the best performance for nine types of behind wall wire corrupted human motions, which is a dramatical improvement when compared to the other state-of-the-art de-wiring methods. The processing speeds of different de-wiring methods are shown in

Table 7. Processing speeds and classification results of different de-wiring methods.

Method	FPS	Accuracy
None	—	$0.934$
Gc	1036	$0.972$
CA-CFAR	3.07	$0.847$
cGAN	14.1	$0.882$
Orignal Cycle GAN	8.50	$0.855$
proposed	8.13	$0.993$

, which is indicated by frames per second (FPS). Gc shares the shortest processing time for each spectrogram via simple exponentiation transform. On the contrary, the CA-CFAR suffers the worst fps because of the time-consuming noise extraction step. Adopting the deeper resnet [19] as the backbone architecture, the proposed Cycle GAN can not accomplish the spectrogram de-wiring task as fast as the cGAN. However, the proposed Cycle GAN can still cover the requirement of in-time de-wiring with a processing speed of 8.13 FPS. In addition, the average classification accuracy for each method is listed below in Table 7. The accuracy is calculated as:

$$\mathrm{accuracy}=\frac{\mathrm{TP}}{N_{data}}$$

(8)

where $\mathrm{TP}$ means the number of true predictions of each de-wiring method, and $N_{data}$ represents the total number of the samples. Obviously, the proposed de-wiring Cycle GAN achieves the highest average classification accuracy $99.3\%$, which further demonstrates that the proposed method precedes the others in de-wiring performance.

5. Conclusions

In this paper, we utilize Cycle GAN for the purpose of spectrograms de-wiring in TW detection scenario. The enhanced micro-Doppler signature is collected in through-the-wall experiment with or without live wire buried in the wall. We choose a Cycle GAN based architecture to deal with the unpaired spectrogram-translation task. The identity loss is integrated into the objective to ensure better realistic spectrogram reconstruction. In the de-wiring part, the proposed method only adopts the forward mapping function. The comparisons of the reconstruction and classification results between the proposed Cycle GAN and other state-of-the-art de-wiring methods indicate the overall advantages of the former in terms of reconstructing the human motion micro-Doppler signatures, and further verify the validity of the proposed architecture for the de-wiring task with the unpaired training samples.

In the future, a better live wire modeling method will be presented for fair comparisons between the cGAN and Cycle GAN. What’s more, the de-wiring spectrogram will be tested by suitable lightweight classifiers for further evaluation.