ZHU Zhengliang, YANG Degui, ZHANG Junchao, and TONG Feng

1.Key Laboratory of Underwater Acoustic Communication and Marine Information Technology of the Ministry of Education, Xiamen University, Xiamen 361005, China; 2.School of Aeronautics and Astronautics, Central South University, Changsha 410083,China; 3.College of Ocean and Earth Sciences, Xiamen University, Xiamen 361005, China

Abstract: Ultra-wideband (UWB) through-wall radar has a wide range of applications in non-contact human information detection and monitoring.With the integration of machine learning technology, its potential prospects include the physiological monitoring of patients in the hospital environment and the daily monitoring at home.Although many target detection methods of UWB through-wall radar based on machine learning have been proposed, there is a lack of an opensource dataset to evaluate the performance of the algorithm.This published dataset is measured by impulse radio UWB (IR-UWB) through-wall radar system.Three test subjects are measured in different environments and several defined motion status.Using the presented dataset, we propose a human-motion-status recognition method using a convolutional neural network (CNN), and the detailed dataset partition method and the recognition process flow are given.On the well-trained network, the recognition accuracy of testing data for three kinds of motion status is higher than 99.7%.The dataset presented in this paper considers a simple environment.Therefore, we call on all organizations in the UWB radar field to cooperate to build opensource datasets to further promote the development of UWB through-wall radar.

Keywords: impulse radio ultra-wideband (IR-UWB), throughwall radar, human motion status, dataset, convolutional neural network (CNN).

1.Introduction

Ultra-wideband (UWB) through-wall radar is widely used in civil and military fields because of its strong penetrability to non-metal and non-transparent obstacles and high range resolution.Impulse radio UWB (IR-UWB)through-wall radar, as a form of UWB through-wall radar, has relatively simple hardware structure.The transmitted impulse has a wide frequency range and it can overcome the influence of radar blind areas.Its applications mainly include post-disaster rescue [1] (e.g., earthquakes, mudslides, mine disasters), anti-terrorism rescue[2,3], and non-contact life monitoring [4,5].As a common non-contact detection technology, radar has advantages in environmental applicability and privacy protection compared with visible light, infrared, and acoustic technologies [6]; it has broad application prospects in indoor human monitoring.In recent years, with the improvement of computing power and the development of artificial intelligence technology, machine learning has shown great promises in the application of UWB throughwall radar target recognition.

Aiming at the problem of the wrong rescue caused by common animal breath frequency range similar to humans in earthquake rescue, the recognition of humans and dogs based on UWB radar was studied in [7], and a support vector machine (SVM) method based on recursive feature elimination was proposed.The proposed method was designed to optimize the radar echo features of humans and dogs to increase the performance of the recognition model.In the through-wall condition, the method proposed in [7] obtained excellent performance to distinguish humans and dogs.In [8], the distinguished method between humans and common animals (e.g., cats, dogs,rabbits) based on the multiscale residual attention network was proposed.The method relied on the difference in the respiratory signal spectrum between humans and common animals.Therefore, the input data size was analyzed from the perspective of network design and predicted performance.If the input data dimension is large,the calculation amount of the distinguished network increases; while if the input data dimension is small, it is difficult to reflect the breath information of the human and it will damage the distinguished accuracy.Literature[9−11] carried out related work for the recognition of human motion status based on carrier-free UWB radar named SIR-20.A variety of feature extraction methods for radar received signals were proposed, such as 2Dvariational mode decomposition (VMD) method, and 2Ddiscrete wavelet transform (DWT) method.In [12], the radar signals of eight human volunteers performing eight different status (e.g.walking, running, rotating, punching,jumping, transitioning between standing and sitting,crawling, standing still) were collected using on UWB through-wall radar.The principal component analysis(PCA) was used to compress the signals and the PCA coefficients were adopted to describe the feature of different human status.Combined with the SVM classifier,the recognition accuracy of different status is higher than 85%.

In [13], the recognition method of human finer-grained activities under the through-wall condition was discussed based on the stepped frequency continuous wave (SFCW)UWB through-wall radar.A comprehensive range accumulation time-frequency transform (CRATFR) based on inverse weight coefficients was proposed to enhance the micro-Doppler features of human finer-grained activity signals.Based on the selected features in the CRATFR spectrum, SVM was adopted to classify different activities.In the experiment with a fixed position behind the wall, the classification accuracy of the six activities performed by eight volunteers was higher than 93.23%.When the human position is changed, the recognition accuracy of the five activities with six meters behind the wall is 86.67%.Reference [14] proposed a fine-grained micro-Doppler feature extraction method for human activities in through-wall environment based on multiple Hilbert-Huang transforms (MHHT).Combined with the principle of human kinematics, it is proved that there is a connection between the MHHT features components and human body structure.A segmented convolutional gated recurrent neural network was adopted to recognize human activity measured by UWB radar in [15].Unlike some methods which used the micro-Doppler spectrograms as input data to train the recognition model, the proposed method extracts the segmented features of micro-Doppler spectrum through convolution operation and uses gate recursive unit to encode the feature maps along the time dimension.The experimental results show that the method can deal with human activities in any length and has a better performance in fine temporal resolution,noise robustness, and generalization.

Compared with the computer vision domain (e.g.ImageNet, the dataset includes 1.5 million images [16]),radar human recognition based on machine learning lacks a large radar target dataset.Therefore, related scholars have explored the research of human status recognition methods based on transfer learning [17,18] and small sample learning [19,20].In [18], the human status recognition method was proposed based on unsupervised domain adaptation.To solve the problem of the insufficient training data in actual measurement, the motion capture dataset (MOCAP) [21] is used to pre-train the recognition network, and via the domain adaptation method combined with measured data to achieve human status recognition.To overcome the problem of unbalanced data, the adaptive incremental recursive least-squares regression parameter estimation method using the UWB radar network was proposed in [19].The proposed method adaptively segments the dynamic signals and realizes the recognition of the human motion status under a small sample condition via the tensor deep learning model.The accuracy of the recognition of the four-motion status under the condition of three radars networked is 91.62%.

From the previous literature review, we can see that a lot of research work focuses on feature extraction, design,and application of machine learning models.However,there is still a problem in the recognition of human status based on IR-UWB through-wall radar, that is, there is a lack of public human motion status radar dataset, and it is difficult to compare the performance of the existing recognition algorithm on the same baseline.In this paper, a variety of human motion status radar echoes under different penetrating media are collected using the IR-UWB through-wall radar system.Then, the IR-UWB human motion status dataset is constructed after preprocessing.The preprocessing algorithm was introduced in [22].The dataset has been uploaded to GitHub (https://github.com/ZhengliangZhu-2020/IR-UWB-Through-wall-Radar-Human-Motion-Status-Dataset).Finally, to help other scholars to conduct related research based on this dataset, we propose a method for recognizing human motion status based on convolutional neural networks (CNN) for reference.

The paper is organized as follows.In Section 2, the human motion model of the UWB radar is briefly introduced and simulated.Based on the feature description of the simulation echoes, the motion status of humans in the measured scene is defined.In Section 3, the human motion status data in different human bodies and different penetrating media is collected based on IR-UWB throughwall radar system, converted into images and constructed as a dataset.In Section 4, a usage demo is provided based on the dataset using CNN and the specific recognition process is given.Finally, conclusions and prospects are provided in Section 5.

2.Theory and model

2.1 UWB radar human motion model

Since the IR-UWB through-wall radar has a high range resolution, the emitted electromagnetic wave is reflected by the human target and then received by the receiver to form an echo.The human return signal is composed of the superposition of different time delays of human multiple scattering centers (e.g.arm, limb, main torso) and the surrounding environment (e.g.ground, windows,roof).These scatterings have different scattering coefficients [23].Then, the UWB radar received signal can be expressed as

wherep(t) denotes the transmit signal of IR-UWB through-wall radar,r(t) denotes the received signal,Nis the number of multipath reflection from human scattering centers and the surrounding environment,Anand τnare the amplitudes and time delay of the corresponding scattering center of the human and surrounding environment, respectively.In a short observation time, the received echo signal can be used to characterize the motion status of humans, the reason is that the different status will cause the parameters in (1) to change.An improved UWB signal model was introduced in [24], in the following we extend the model to simulate UWB human motion received signals.In this simulation, we set the following assumptions: the signal transmitted by the radar does not attenuate during propagation, and the transmitted signal is the Gaussian pulse signal; the transceiver antenna of the radar is in ideal conditions; the scattering center of human-only considers the torso and the displacement of chest; the human motion status is walking and standing still.

The human motion status is shown in Fig.1.In Fig.1(a), consider the human walking with two trajectories.The first trajectory is walking back and forth along the radar line of sight radially, and its round-trip period is 12 s.The second trajectory is walking perpendicular to the radar light of sight tangentially, and its round-trip period is 10 s.In Fig.1(b), with the human standing still,the human simulation respiratory frequency is 0.3 Hz and the corresponding chest placement is 0.05 m.

Fig.1 Schematic diagram of simulation scene

Then, the simulation receive signals can be expressed

where, c=3×108m/s is the propagation speed of electromagnetic waves, τd(t) is the time delay as the human trajectory changes,d(t) is the human trajectory with a function of timet.Discrete (2) along the fast time dimension (i.e., range dimension) and slow time dimension, we can obtain the element of the simulation receive signal matrixR:

wheremandTsrepresent the sampling points and sampling interval along the slow time direction, respectively;nandTfrepresent the sampling points and sampling interval along the fast time direction, respectively;Apdenotes the amplitude of the transmit signal; ρ denotes the pulse width parameter of the transmit signal.

Fig.2 shows the corresponding simulation results.Generally, it is called the time-domain range profile,which characterizes the distance relationship between the human and radar.In other words, the time-domain range profile can be used as the basis for recognizing human motion status.

Fig.2 Simulation results of UWB radar human motion model

2.2 Human target echo signals feature description and motion status definition

According to the research of target motion status recognition based on UWB radar received signals, the different feature description methods can be divided into time-domain range and time-frequency domain based ones.The main purpose is to analyze the radar echoes characteristics in a short-term observation time window to recognize the human motion status.

Time-frequency domain feature description: analyze radar echoes characteristic via the time-frequency transformation methods, e.g., short-time Fourier transform(STFT), Wigner-Ville distribution (WVD), wavelet transform (WT), Hilbert-Huang transform (HHT).Youngwook et al.[25] used Doppler radar to collect the echoes of human different activity and transformed the signals into the time-frequency domain for feature extraction via STFT.The extracted features and definitions are as follows:

(i) Torso frequency reflects the speed of a human, the speed of the human torso varies greatly depending on the human motion status;

(ii) Total bandwidth (BW) of the Doppler signal describes the speed of human limb, e.g., arms/legs produce a larger Doppler signal BW;

(iii) Offset of the total Doppler signal represents the asymmetry of human limbs between moving forward and backward;

(iv) BW without micro-Dopplers denotes the swing changes of human torso;

(v) Period of the limb motion describes the swing rate of the arm or the leg.

Fig.3 shows the illustration of the features.

Fig.3 Features illustration of the time-frequency spectrum [25]

Time-domain range profile feature description: take the result of Fig.2 as an example, and the features extracted are shown in Fig.4.The specific features are defined as follows:

Fig.4 Features illustration of the time-domain range profile

(i) Period describes the time needed for the human to complete the walking trajectory.When the human stands still, this value can be approximated as the respiratory cycle.

(ii) Chest motion range denotes slight displacements of the chest cavity due to respiration.

(iii) Human motion range reflects the max distance between the radar and the human.

(iv) Instantaneous velocity angle describes the instantaneous velocity of the human, and the greater the human velocity is, the greater the value will be.

To realize the penetration performance of non-metallic obstacles, IR-UWB through-wall radar has a low center frequency and is limited in spatial resolution and micro-Doppler sensitivity, and it has a weak ability to acquire human body movement information, so it can only perform rough recognition of human motion status.It should be noted that the IR-UWB through-wall radar is a timedomain radar, so it is hard to take the micro-Doppler effect into consideration.Therefore, in this paper, human motion status is defined as follows:

(i) Human walking: human moves back and forth in radial and tangential direction within the power of radar detected beam;

(ii) Human standing still: human stands still within the power of the radar detected beam.Except for the fretting caused by normal breathing and heartbeat, the human is accompanied by some small-amplitude oscillation.

Considering that the human motion status occurs in the detection scene, then the empty scene (i.e., the human target does not exist in the detection scene) is defined as the human motion status.

3.Radar system and data acquisition

3.1 IR-UWB through-wall radar system

IR-UWB through-wall radar used in this paper is independently designed by the School of Aeronautics and Astronautics, Central South University, China.The structure of the radar is shown in Fig.5, and the specific design method of its submodule can be referred to[26,27].The improved butterfly antennas are used in this radar system and it can be operated wirelessly with a tablet terminal or with a cable connected to a personal computer.Table 1 shows the key parameters of the IR-UWB radar.The pulse signal generating module generates a trigger pulse with a pulse repetition frequency of 400 kHz, and the center frequency of the transmitted signal is 500 MHz.After the transmitted signal is scattered by the target, the return signals are sampled by the echo acquisition module and processed by the field programmable gate array (FPGA) for the further step.Due to the narrow pulse width of the transmitted signal, it is necessary to use an analog-to-digital converter (ADC) with a high realtime sampling rate for sampling but it may lead to a high cost.Therefore, we use the equivalent sampling technology in the design of the echo acquisition module.The real-time sampling frequency of the ADC is 400 kSa/s and using this method, the sampling frequency can reach 5 GSa/s.The sample number of the received signals is 768 in the fast time direction.In the slow time direction,eight return signals are received in 1 s.To improve the signal-to-noise-ratio (SNR) of the signal, the average processing of multiple sampling at a single point is carried out.Therefore, the record time of a single frame echo signal can be calculated as

Fig.5 Structure of IR-UWB through-wall radar

Table 1 Parameters of IR-UWB through-wall radar

3.2 Time-domain range profile date acquisition

The acquisition of time-domain range profile data is obtained by the actual measurement of the human motion status using IR-UWB through-wall radar, and the corresponding experimental scenes are shown in Fig.6.Consider the wall-effect on the radar return signal has been deeply studied in [28].The conclusion is that the physical parameters of the wall (especially the dielectric constant and thickness of the wall) will cause the human positioning deviation but it has little effect on human motion status.The IR-UWB radar is close to the wall, and the height of radar is 1.5 m from the ground.The dielectric material penetrated is a brick-concrete wall with a thickness of 0.2 m.In the experimental setup, the human walking range is 3 m to 8 m.The volunteers participating in the experiment are three healthy adults.Based on the above experimental conditions, the dataset contains a total of 4230 time-domain range profile data with 768×32 pixel and the observation time of each sample is 4 s.

Fig.6 Experimental scenes

Table 2 and Table 3 are the time-domain range profile data of different humans at different times and in different motion status.By analyzing the time-domain range profile data sample from these tables, the following conclusions can be obtained:

Table 2 Time-domain range profiles of Volunteer NO.2 in different motion status under different penetrating media

Table 3 Time-domain range profiles of Volunteer NO.3 in different motion status under different penetrating media

(i) When the human is in different motion status, the corresponding time-domain range profile is different,which is roughly the same as the time-domain range profile obtained by simulation.

(ii) When the human is in the same motion status, the corresponding time-domain range profile features are consistent.In other words, in the same motion status, the time-domain range profile features are invariant and independent of a human target.

(iii) When penetrating different media, the time-domain range profile is still invariant, because the different media mainly cause the distortion of electromagnetic wave propagation speed, amplitude, and other parameters, which finally affects the accurate positioning of the human.

(iv) The longer the observation time is, the more sufficient the description of the human motion status is, which is more conducive to the recognition of the human motion status, but it will affect the recognition efficiency.

In conclusion, the construction of the IR-UWB throughwall radar human motion status dataset is completed.Fig.7 shows the composition of the human motion status dataset.The construction of the dataset provides data support for the realization of the subsequent human motion status recognition algorithm.In the following section in this paper, we introduce a human motion status recognition algorithm.

Fig.7 Composition of IR-UWB through-wall radar human motion status dataset

4.Demo: human motion status recognition using CNN for IR-UWB through-wall radar

Based on the dataset collected in Section 3, in this section, we introduce how to use the dataset to recognize the human motion status.This section mainly provides a dataset division ration from the perspective of machine learning.Based on the divided dataset, CNN is trained to realize the recognition of human motion status.

4.1 Dataset division

The constructed dataset needs to be divided.Generally,the dataset is divided into a training set and a test set.The training set is used to train the model to learn the task and adjust the parameters of the model.The test set is used to evaluate the performance of the trained model.In short,the training set enables the model to abstract general rules, while the test set is used to check whether the rules abstracted from the model are correct.Generally speaking, the following suggestions should be considered in the division of the dataset:

(i) The selection of the test set and the training set should be mutually exclusive as far as possible [29], and the sample of the test set should not be used in the training set.If the sample of the test set is used as the sample of the training set, the evaluation result of the whole model will be too optimistic.

(ii) The division of the test set and the training set should keep the proportion of data categories similar.Taking the dataset constructed in Section 3 as an example, when dividing the training set, the number of samples corresponding to each human motion status should be equal as possible.If the sample imbalance occurs in the training set, the model will tend to the class with more samples, which is not conducive to the generalization of the model.

(iii) The number of samples in the training set and data set should maintain a certain proportion, this suggestion is considered from the overall dataset.Reference [29]mentioned that when the dataset is divided into a training set and a test set, it will face the dilemma of deviationvariance.Therefore, 2/3 − 4/5 of the dataset is used as the training set, and the rest data is used as the test set.

(iv) The division should consider the training set and test set samples from the same distribution.Besides, the training set also needs to be divided into a part of the validation set to adjust the supper parameters of the trained model.

To sum up, the IR-UWB human motion status dataset is divided as follows: the ratio of the training set, validation set, and test set is 6∶2∶2, as shown in Fig.8.The corresponding sample numbers are 2592, 864, and 864.

Fig.8 Division of IR-UWB through-wall radar human motion status dataset

4.2 Human motion status recognition based on CNN

4.2.1 CNN theory and design

CNN [30,31] is one of the deep learning algorithms widely used in computer vision.Driven by large data, the recognition or classification results are directly output via end-to-end deep CNN.Fig.9 is the CNN used in target recognition by Doppler radar [32], and MLP means multilayer perceptron.

Fig.9 CNN used in Doppler radar target recognition in [32]

The essence of CNN is a hierarchical model and its infrastructure is the combination of a convolutional layer, a pooling layer and a full connected layer.By adding the convolutional layer, and the pooling layer, the network depth is deepened and more local features are obtained.This framework is based on the conclusion of Hubel’s research on the neurons of the cat’s visual cortex [33]: in the process of human visual brain cognition, low-level specific features are continuously transferred to highlevel abstract features via neurons.CNN uses multiple convolutional layers and pooling layers to simulate the human visual brain cognitive mode and obtain local features.Then the local features are integrated into the form of a full connected layer.In the convolutional layer, the local features are obtained by the inner product operation of the convolutional kernel and input data, and the nonlinear transformation of the activation function.Here are some common activation functions: Sigmoid function,Than function, and Relu function.Convolutional kernels with different sizes can be interpreted as feature extraction with different scales for input data.The zero-padding operation is often performed to ensure the size consistency of input data.The main purpose of the pooling layer is to reduce the dimension of feature data and restrain the overfitting of the network.The pooling layer reduces the number of parameters of the trained model and improves the operational efficiency of the network model.The common pooling methods are Max pooling, Average pooling.In the full connected layer, the softmax function is used to solve the multi-classification problem.Assuming that there arekclassification problems, the output of the softmax function can be calculated as

whereW=[W1,W2,···,Wk] andb=[b1,b2,···,bk] are the corresponding weights and biases.The training process of CNN includes two parts: forward propagation and error back-propagation.The training goal is to minimize the loss function of the whole network.The error back-propagation updating model parameters are driven by error.The methods like SGD, Momentum, Adam, and RmsProp are proposed to improve the update speed of parameters.

In this paper, we design the CNN for IR-UWB throughwall radar human motion status recognition.Table 4 is the specific parameter setup of CNN.The design inspiration of the CNN comes from the structure form of VGGNet [34], namely the combination of multiple convolutional layers and pooling layer, and the number of the convolutional kernel in each layer shows a certain change rule, as shown in Fig.10.The floating point operations(Flops) is used to measure the computational complexity of the proposed CNN, as shown in Fig.11.Meanwhile,the parameters number of each layer is also shown in Table 4.

Table 4 Parameters setup of CNN

Fig.10 Structure comparison between the VGGNet and CNN proposed in this paper

Fig.11 Flops of the proposed CNN

4.2.2 Process of human target motion status recognition and results analysis

According to the description of the dataset divided and the principle of CNN in Section 4, the flowchart of human motion status recognition based on CNN is as follows:

Step 1Read the training set samples of the dataset,get the time-domain range profile and its corresponding label, and normalize the time-domain range profile globally.

Step 2Using the time-domain range profile as the input sample of the CNN.Set up the parameters of CNN such as activation function type, iteration times, and network structure.In the training process, CNN automatically extracts the features of the time-domain range profile.Finally, the output of the network output layer is used to calculate the error between the predicted output and the actual output, and the error back-propagation to update the parameters.The weight and bias of the network are updated by using the accelerated gradient descent algorithm (RMSprop).This is repeated until the cost function meets the iterative conditions, and the training is completed to get the appropriate target operation.

Step 3Fix and save the trained CNN model of human motion statue recognition.

Step 4Read the test set samples of the dataset, process them as described in Step 1, and input them into the well-trained CNN in Step 3.Then recognition of the human motion status can be obtained.

Fig.12 is the flow chart of the target motion status recognition algorithm based on CNN.The CNN is implemented with Keras running on top of Tensorflow.The CNN is performed on the computer with 3.6 GHz Intel(R) Core (TM) i9-9900K CPU (64G RAM) and GeForce RTX 2080Ti GPU.In the training process of recognition model, the training is performed for 50 epochs with a batch size of 100, the learning rate is 0.001.We calculate and record categorical cross-entropy as the loss function in the training stage.Fig.13 shows the recognition accuracy and categorical cross-entropy value of the training set.It can be seen from the figure that with the increase of iteration times, the accuracy of the training set gradually tends to be stable and the cross-entropy loss tends to converge.

Fig.12 Flowchart of human motion status recognition based on CNN

Fig.13 Recognition accuracy and cross-entropy loss of training set data

To verify the feature extraction ability of the training network, the t-distributed stochastic neighbor embedding(t-SNE) algorithm [35] is used to map the high-dimensional features extracted by the trained CNN set to twodimensional for feature visualization.As shown in Fig.14, it can be seen that the trained CNN can well extract the features of three kinds of motion status of human targets.Table 5 shows the performance results of the trained CNN in the test set.The first row in the table represents the actual motion status of the sample, and rows 2 to 4 represent the predicted status of the network.The recognition accuracy of the training model is 99.7%.Especially, 100% accuracy can be achieved for human walking and empty scene.

Table 5 Confusion matrix of the test set data

Fig.14 Features of t-SNE mapping method

For reflecting the benefits of the CNN, using a similar method proposed in [20], a comparison is made with the stacked auto encoder (SAE)-SVM algorithms with different parameters (the SAE is used to extract features for the input data, and the SVM is employed for make a classification, the detailed parameters of these methods are shown in Table 6).In Table 6, Input [x] means input layer; H[x] means hidden layer; F[x] means feature layer;xis an integer, meaning the number of neurons; the SAE uses the Relu function as an activation function for each layer and uses the Adam as the parameter update method.Fig.15 shows the flow chart of this comparison method.In the same test dataset, the performance of these algorithms is worse than the proposed CNN, as Fig.16 shows.

Table 6 Parameters of the SAE and SVM

Fig.15 Flowchart of human motion status recognition based on SAE+SVM

Fig.16 Comparison result between CNN and SAE-SVM

5.Conclusions and future prospects

In this paper, the dataset of human motion status based on the IR-UWB through-wall radar is published, hoping to promote the further integration of the UWB through-wall radar human target recognition and machine learning algorithm.Based on the human motion model of UWB radar, the time-domain range profile simulation of humans in different motion status is performed.Using the simulation results, the time-domain range profile features under different motion status is analyzed and used to guide the human target motion status in the actual experimental situation, namely human walking, human standing still and empty.For the published datasets, this paper proposes a CNN-based method for the recognition of human motion status.By using the trained CNN, the recognition accuracy of human motion status can reach 99%.Our results show that machine learning has potential in human target recognition of through-wall radar.Although the provided cases have a good recognition performance, we need to realize that the method based on CNN may have high computational complexity.Therefore, in practical application deployment, we should focus on the complexity and timeliness of the CNN model.

In future research, it is suggested that the work of human target recognition of UWB through-wall radar should be deeply explored from the following directions:

(i) In the dataset proposed in this paper, because the definition of human motion status is relatively simple, the feature description of time-domain range profile is adopted, and this method depends on the time length.However, if the selected time length is too long, two motion status may overlap in the same observation window and reduce the recognition efficiency.Therefore, it is necessary to explore the method of human target motion status recognition based on the instantaneous features,which can fundamentally remove the dependence on time length and can estimate the duration of motion status.This kind of method is real-time and will be helpful to further study the fine recognition of human motion status.

(ii) At present, most of UWB through-wall radar work in low frequency are limited in spatial resolution and Doppler sensitivity to ensure their detection ability in nonmetallic and non-transparent media attenuation, so it is difficult to extract the spatial information of human limbs.Therefore, it is necessary to improve the ability of human information acquisition at low frequency.Combined with the human motion scattering model [36,37],the robust feature representation method for describing fine human motion state should be further studied.Meanwhile, we also need to focus on the multi-dimensional feature extraction related to humans and study the identification algorithm of human target based on the human motion state recognition.

(iii) Although the research of the UWB through-wall radar system and its algorithm were stated in 1980s and has made great progress, there is a lack of the UWB through-wall radar dataset for other scholars to study, and the recognition algorithm based on machine learning often needs sufficient data to complete the algorithm development.Therefore, we call on relevant organizations to cooperate in UWB radar data acquisition in complex scenes,to promote the further development of UWB throughwall radar human target recognition.