Dense People Counting Using IR-UWB Radar with a Hybrid Feature Extraction Method

1 Dense People Counting Using IR-UWB Radar with a Hybrid Feature Extraction Method Xiuzhu Y ang, W enfeng Y in, Lei Li and Lin Zhang Beijing Uni versity of Posts and T elecommunications Email: zhanglin@bupt.edu.cn Abstract —People counting is one of the hottest issues in sensing applications. The impulse radio ultra-wideband (IR- UWB) radar has been extensively applied to count people, pro viding a de vice-free solution without illumination and pri vacy concerns. However , perf ormance of current solutions is limited in congested en vironments due to the superposition and obstruction of signals. In this letter , a hybrid feature extraction method based on curvelet transf orm and distance bin is proposed. 2-D radar matrix features ar e extracted in multiple scales and multiple angles by applying the curvelet transf orm. Furthermore, the distance bin is introduced by dividing each row of the matrix into several bins along the pr opagating distance to select features. The radar signal dataset in three dense scenarios is constructed, including people randomly walking in the constrained area with densities of 3 and 4 persons per square meter , and queueing with an av erage distance of 10 centimeters. The number of people is up to 20 in the dataset. Four classifiers including decision tr ee, AdaBoost, random f orest and neural network are compared to validate the hybrid features, and random fores t performs the highest accuracies of all above 97% in three dense scenarios. Moreo ver , to ensure the reliability of the hybrid featur es, thr ee other features including cluster features, activity features and CNN features are compared. The experimental results re veal that the proposed hybrid feature extraction method exhibits stable performance with significantly superior effectiveness. Index T erms —People counting, IR-UWB radar , hybrid featur e extraction, curvelet transform, distance bin, random for est. I . I N T RO D U C T I O N W ith the developing requirement for the Internet of Things (IoT) sensing task, estimating the number of people in a mon- itored area is crucial for sensing applications. Radar systems provide device-free sensing solutions ranging from human detection to acti vity classification [1], [2]. They leverage radar signals which are reflected and attenuated by human bodies, and infer the v alid information by properly analyzing the receiv ed signal. The impulse radio ultra-wideband (IR-UWB) radar transmits and receives a narro w impulse signal that occupies a wide bandwidth in the frequency domain, with fine delay resolution and excellent penetration. It performs outstanding applications in vital sign monitoring [3], personnel detection [4] and people counting [5]-[7]. Compared with current researches on people counting using vision-based systems [8], the IR-UWB radar doesn’t suf fer from insufficient illumination and priv ac y concerns. Moreo ver , it is a de vice-free solution without relying on any dedicated or personal de vice, which is required in other radio-based systems, such as radio Corr esponding author: Lin Zhang frequency identification (RFID), Bluetooth, Zigbee and W iFi [9]. Sev eral studies on people counting using IR-UWB radar are conducted in [5]-[7]. The algorithm in [5] iterati vely detects the local maximum of radar signals to count people. In [6], theoretical models of UWB signals are conducted in simulation. [7] proposes an algorithm based on the major clusters, analyzing the distribution of selected amplitudes with the distance and the number of people. These algorithms adequately distinguish multipaths and count people. Ho wev er , all of them count each signal separately suf fering from e ver - changing signals, and the superposition as well as obstruction of signals limit counting performance in congested en viron- ments. In this letter , a hybrid curvelet transform based features- distance bin based features (CTF-DBF) extraction method for dense people counting is proposed. Firstly , in order to address challenges of rapid variations between signals and superposed multipaths of each signal in congested scenarios, sev eral continuously recei ved signals are regarded as a 2-D radar matrix. Due to the moving continuity and trajectory consistency of people, characteristics of moving people are represented as textures with spatial locality information in the radar matrix. The curvelet transform is applied to extract statistical features in multiple scales with different frequencies as well as multiple angles with di verse moving directions. Secondly , to extract detailed information and further analyze the superposed and obstructed signal, the distance bin is defined by dividing each signal into sev eral bins along the propagating distance. Characteristics of each distance bin are extracted in an effecti ve way to supplement detailed features for statistical features. The radar signal dataset comprising three dense scenarios is constructed for 0-20 people randomly walking in the constrained area with densities of 3 and 4 persons per square meter, and at most 15 people queueing with an average distance of 10 centimeters. W ith these hybrid features extracted from the dataset, four classifiers including decision tree, AdaBoost, random forest and neural network are compared. Random forest achie ves the highest accuracies in three dense scenarios of all abov e 97%. Furthermore, three other features including cluster features proposed in [7], activity features in [10] and features learnt automatically form LeNet-5 con volutional neural network (CNN) [11] are compared to ensure the reliability of the hybrid features. The experimental results demonstrate the effecti veness and robustness of the hybrid feature extraction method in dense 2 R e c e i v e d R a d a r D a t a ( 5 0 x 1 2 8 0 ) S u b - b a n d 1 C o a r se La y e r S u b - b a n d 2 D e t a i l L a y e r S u b - b a n d 3 F i n e La y e r R e c e i v i n g t i m e D C R e mo v a l a n d B a n d P a s s F i l t e r i n g Cu r ve l e t t r an sfor m b ase d f e at u r e s S e l e c t i n g c o e f f i c i e n t s i n d i f f e r e n t d i r e c t i o n s Distan c e b i n b ase d f e at u r e s …… 1 x 1 2 8 0 d i s t a n c e b i n R a n d o m F o r e s t …… C l u t t e r R e m o v a l C u r v e l e t D o ma i n e n e r g y a n d me a n v a l u e o f c u r v e l e t c o e f f i c i e n t s S e l e c t i n g d i r e c t i o n s w i t h ma i n e n e r g y e n e r g y r e c o n s t r u c t i o n e n e r g y a n d t o p f i v e ma x i mu m v a l u e s o f c u r v e l e t c o e f f i c i e n t s Re f i n e d D at a A m p l i t u d e D i s t a n c e i n d e x ( s a m p l e s ) a v e r a g e d f e a t u r e s w i t h e a c h d i s t a n c e b i n C l a s s i f y i n g t h e n u mb e r o f p e o p l e B an d p ass D at a D i s t a n c e i n d e x ( s a m p l e s ) B a ndpa s s S ig na l R e f in e d S ig na l Da ta se t G e n e r a tio n i n p u t i n p u t H ybri d F eat ures ( 1 x 300 ) Class ific at io n H y b r id Fea tu r e Ex tr a c tio n Me th o d S ig n a l Pr e p r o c e s sin g Fig. 1: W orkflo w of the people counting system, composed by the dataset generation module, the signal preprocessing module, the proposed hybrid feature extraction method module and the classification module. scenarios. Fig. 1 shows the workflow of the people counting system, composed by the dataset generation module, the signal prepro- cessing module, the proposed hybrid feature extraction method module and the classification module. The remainder of this paper is organized as follo ws. Section II describes the dataset generation. The proposed hybrid feature extraction method is discussed in Section III. Section IV presents experimental re- sults and analysis. The conclusions are summarized in Section V . I I . D AT A S E T G E N E R A T I O N A. Radar System In this letter, the IR-UWB radar data from a select number of people in a space is acquired by an NV A-R661 radar module transmitting a narrow pulse with a center frequency of 6.8 GHz, and the bandwidth in -10dB concept of 2.3 GHz. The receiv ed radar signals are conv erted to digital signals, and the sampling frequency is about 39 GHz with the resolution of 0.0039 meter . The experimental setup is sho wn in Fig. 2, in which the radar is installed at a height of 1.8 meters, with the detecting range of 5 meters and a central angle of 90 degrees. T o validate the performance of the proposed method, three dense scenarios are considered for radar data collection. Scenarios 1 and 2 are 0-20 people randomly walking in a constrained area with densities of 3 and 4 persons per square meter respectively . T o maintain the densities unchanged, the activity range of testers is limited in a rectangular region of which the area increases with the increasing number of people, sho wn in the red area in Fig. 2(a). Due to the congested en vironment, the moving speed of people is limited, equal to or less than the normal walking speed. In scenario 3, at most 15 people stand in a queue with an average distance of 10 centimeters described in Fig. 2(b), and their positions are unchanged. T o enhance the reliability of the dataset, 44 testers participate in e xperiments for acquiring diverse data from dif ferent people. Fiv e seconds of radar data with 200 receiv ed signals are recorded for each measurement, and each radar sample is se- lected independently from the record for 1.25 seconds with 50 receiv ed signals. Radar sample for 1.25 seconds is long enough for curvelet transform based feature extraction, meanwhile counting in every 1.25 seconds is acceptable in a real time system. Each signal in a radar sample contains 1280 sampling points representing the 5 meters detection range. 3,360 radar samples are generated in scenarios 1 and 2 respectiv ely , with a total of 2,560 samples in scenario 3. ( a ) 5 m 90 ° Rad ar 1 . 8 m 5 m 1 . 8 m R ad ar 5 m ( b ) 1 . 7 m Fig. 2: Experimental setup (a) in the constrained area and (b) in a queue. B. Signal Modeling For each receiv ed radar sample, the direct current (DC) component is firstly remov ed, and then a Hamming window is designed as a filter to obtain the bandpass data with frequency from 5.65 GHz to 7.95 GHz, shown in Fig. 1. The 2-D bandpass data composed by multiple radar signals is described as follows, r ( t, x ) = p ( t, x ) + c ( t, x ) + n ( t, x ) (1) where t is the accumulating receiving time representing the time it takes the radar to recei ve multiple signals, while x is the propagating distance of each signal. p(t,x) is the target signal reflected from people, while n(t,x) is the noise signal. c(t,x) 3 represents the clutter , which contains the direct wave from the transmitter to receiver and reflections from the background. I I I . H Y B R I D F E A T U R E E X T R A C T I O N M E T H O D A. Curvelet T ransform based F eatures Fast ever -changing signals and superposed multipaths bring great challenges for dense people counting by amplitudes of each signal. Counting from a single recei ved signal separately is not stable and reliable, therefore the curvelet transform provides statistical features of a radar matrix. Several tem- poral continuously receiv ed signals are considered as a 2- D radar matrix to avoid contingency caused by the single signal. Furthurmore, considering the moving continuity and trajectory consistency , trajectories of people are presented as textures with spatial locality information in the matrix. Superposed multipaths sho w stronger textures and can be ob- viously observed with the curvelet transform. In addition, the curvelet transform provides a multi-scale and multi-orientation decomposition for the 2-D radar matrix to adequately represent texture and edge information with curve-like features [12], providing information on signal strength and moving direction of people. The definition of discrete Curvelet transform is giv en as follows, C ( j, l , k ) = Z ˆ f ( ω ) ˆ U j ( S − 1 θ l ω ) e i dω (2) where j , l and k are the parameters of the scale, the direction and the position. f represents the input radar data in the Cartesian coordinate system. U j is the frequenc y window for each scale j , and S θ l is the shear matrix with orientation θ l defined as S θ l :=  1 0 − tanθ l 1  (3) where superscript T represents the transpose of the matrix. b is defined as b := ( k 1 · 2 − j , k 2 · 2 − j / 2 ) , where the sequence of translation parameters k = ( k 1 , k 2 ) ∈ Z 2 . When people are queueing in a line or remaining still, their positions are unchanged in a period of time, forming straight-like lines in the 2-D image. In this case, signals from people with smaller v ariances are easily to be mistaken as clutters reflected from background, thus removing clutters will eliminate significant information as well. T o fully extract statistical features from radar matrix without losing any useful information, the bandpass data matrix without clutter remov al is decomposed by the curvelet transform, shown in Fig. 1. Each bandpass data with 50 continuously recei ved signals and 1280 sampling points in each signal is decomposed into a coarse layer , a detail layer and a fine layer representing different scales. T o characterize the signal matrix in all scales, features with three layers are extracted in the curvelet domain. The coarse layer formed by low frequency coef ficients shows the general characteristic and the tendency information of the signal matrix, thus the mean value and energy of the coarse layer are extracted to generally describe the radar data. The fine layer contains high frequency coefficients, representing the finer edge information, which is usually represented by the maximum v alue. Therefore the top fi ve maximum values as well as the ener gy of the fine layer are extracted. P 1 P 14 P 9 P 4 P 5 P 6 P 7 P 8 P 12 P 13 P 15 P 16 E ner g y dis t ribut io n ( % ) 2 1 3 4 5 6 7 8 9 10 11 12 13 14 15 16 P a n el in d ex Fig. 3: Energy distribution of 16 coefficient matrices in the detail layer with corresponding panels in the radar matrix. The red, blue and gray dashed lines parallel to the vertical axis are the energy of corresponding panels marked in the same colour . ( a ) ( b ) ( c ) ( d ) distan c e ti me distan c e time distan c e time distan c e time Fig. 4: The radar matrix for (a) the bandpass data, and the reconstructed data by (b) the 90 ◦ vertical coefficients (c) the 45 ◦ diagonal coefficients (d) the 135 ◦ diagonal coefficients of the detail layer . The detail layer with high frequency coefficients is di vided into 16 directions. Panels of the radar matrix in Fig. 3 are arranged in the clockwise direction, and each angular panel occupies 22 . 5 ◦ . Coefficients in each panel represent signals on corresponding moving directions in trajectories of people. In order to increase the stability and reliability of extracted features, the blue dashed line parallel to the horizontal axis in Fig. 3 cuts the energy , and coef ficients with too low energy are removed. Fig. 4 sho ws the reconstructed signals by picking up coef ficients in corresponding directions, which are presented as textures of moving trajectories in the grey- scale maps. Panel 1, 16, 8 and 9 in blue are selected as the 45 ◦ direction sho wn in Fig. 4(c), representing people moving further away from the radar in 1.25 seconds, while panel 4, 5, 12 and 13 in gray are extracted for the 135 ◦ direction in Fig. 4(d) representing people moving closer to the radar during this time. Panel 6, 7, 14 and 15 in red representing the 90 ◦ direction in Fig. 4(b) occupy most of the energy due to the static clutter , as well as reflections from people on the spot. Considering superposed multipaths for stronger textures, the energy for each direction are calculated in the curvelet domain and in reconstructed signals respectiv ely to represent the comprehensive information of people in the corresponding direction. B. Distance Bin based F eatur es T o get detailed features for complementing curvelet trans- form based statistical features and further analyze the super- posed and obstructed signal in dense people counting, several features are extracted from each signal. Clutter signals are remov ed firstly using a running av erage based method [5] to analyze the valid signals reflected from people, and feature 4 extraction is operated on the refined data in red, sho wn in Fig. 1. Due to the high sampling rate in receiv ed radar signals, the redundant information contained in these samples will cause ov er-fitting. T o select the representati ve information for the number of people and reduce ov er-fitting, each signal with 1280 sampling points is divided into se veral bins along the propagating distance, with each length of S d . D i s tan c e ( m ) A mp l itud e 0 1 2 3 4 5 0 . 8 0 . 6 0 . 4 0 . 2 0 - 0 . 2 - 0 . 4 - 0 . 6 - 0 . 8 d is tanc e b in r e f lect e d s ign a l s n o i s e s Fig. 5: Refined radar signal for 4 people in a queue. The maximum amplitude in each distance bin is selected as a feature to represent a candidate point for the presence of people. Howe ver , the number of local maximum amplitudes can not represent the number of people when the y stand closely . As shown in Fig. 5, the first red circle is obviously detected for the presence of 1 person, but there are 2 persons standing closely . In this case, it is impossible to distinguish the number of people from the amplitude due to the multipaths from dif ferent people. But the ener gy of dif ferent people are superposed and distinct significantly in a distance bin, therefore the energy is calculated as the complementary feature by squaring sampled signals and integrating them ov er each bin. Since that the transmitted power of UWB radar is limited and the relatively high noise is accumulated over its wideband. The signal-to-noise ratio (SNR) is lo w , thus considering the energy of each bin also takes noises into calculation and meanwhile amplifies them. In dense queueing counting, the reflected signal from obstructed people is se verely attenuated and comparable with the noise in observation, as Fig. 5 describes. Amplitudes marked by two orange circles are noises in the environment, but their amplitudes are comparable with reflected signals from people, marked in the second and third red circles. Therefore, the noise is removed using the hard threshold analysis of the curvelet transform [13]. Then the energy and maximum amplitude of denoised signals are also extracted as features. The length of each distance bin S d is of great importance in identifying the detailed features. T o discriminate each person from the superposed multipath signals in congested en vironments, S d should be smaller than a certain physical parameter , for example, the height or the shoulder width of a person [7]. In order to obtain suf ficient features in different scales to better describe the detailed information, the distance bin is chosen for 125, 250 and 500 millimeters respectiv ely . In this letter, the spatial resolution of the radar system for the minimum distinguishable distance between two adjacent sampling points is 3.9 millimeters. Therefore, the number of sampling points in each distance bin is set as 32, 64 and 128 respectiv ely . T o maintain the length of the distance bin, radar system with higher spatial resolution needs more sampling points in each distance bin, while radar system with lower resolution needs less sampling points. For the radar matrix with multiple signals, these features are av eraged respecti vely . The distance bin based features are then directly combined with the curvelet transform based features as the hybrid features, which are defined more clearly in T able I. T ABLE I: Hybrid CTF-DBF features T erms Definition Features of coarse layer in the curvelet domain Mean and energy of curvelet coefficients in the coarse layer Features of fine layer in the curvelet domain T op fi ve maximum values and energy of curvelet coefficients in the fine layer Features of detail layer in the curvelet domain Energy of the 90 ◦ vertical coefficients, 45 ◦ diagonal coefficients and 135 ◦ diagonal coefficients Features of detail layer in the reconstructed signal Energy of the reconstructed signal with the 90 ◦ vertical coefficients, 45 ◦ diagonal coefficients and 135 ◦ diagonal coefficients Number of sampling points in a bin S d The number of sampling points in a distance bin, with domain{32, 64, 128} Maximum Amplitude A k / A d The maximum amplitude of each distance bin for signals with and without noises with corresponding S d Energy E k / E d The energy of each distance bin for signals with and without noises, with corresponding S d I V . E X P E R I M E N TA L R E S U LTS A. P erformance on Dif fer ent Classifiers The hybrid CTF-DBF feature samples extracted from the dataset constructed in this letter with size 1 × 300 are used as input for classification. T o verify the ef fectiv eness of the hybrid feature extraction method, four selected classifiers are compared, including decision tree, AdaBoost, random forest and neural network. The decision tree is a rooted tree structure to di vide the cases into two subtrees in each node. In this letter , the random forest classifier constructs 200 decision trees to increase the classification ability . The AdaBoost consists 50 base estimators, with the SAMME.R classification algorithm and a linear loss function. The neural network has three hidden layer with 100, 200 and 100 neurons respectively as well as a ReLU activ ation function, and is optimized by Adam algorithm. The feature samples are divided into the training set and the testing set. 80% randomly chosen samples (2688 samples in scenarios 1 and 2 respecti vely , and 2048 samples in scenario 3) are used as the training set for the supervised training on each classifier , and the remaining 20% feature samples are as the testing set to test the classifier and calculate the classification error . The calculation for each classifier is repeated for 20 times with randomly chosen training data, and the average accuracy , precision, recall and F1 score [14] are computed, shown in T able II, III and IV . As T able II, III and IV describe, the accuracies of random forest and neural network in three dense scenarios are all abo ve 94%, proving the effecti veness and robustness of the hybrid features in dense people counting. Random forest performs the highest accuracies of 97.6% and 97.5% in the constrained area 5 T ABLE II: Classification performance comparison of dif ferent classifiers for 0-20 people randomly walking in the constrained area with 3 persons per square meter . Accuracy Precision Recall F1 Decision Tree 76.7% 87.6% 87.0% 87.3% AdaBoost 80.5% 90.9% 91.9% 91.4% Random F orest 97.6% 97.5% 99.5% 98.5% Neural Network 95.1% 97.8% 95.3% 96.5% T ABLE III: Classification performance comparison of dif- ferent classifiers for 0-20 people randomly walking in the constrained area with 4 persons per square meter . Accuracy Precision Recall F1 Decision Tree 76.6% 87.6% 87.0% 87.3% AdaBoost 81.2% 90.9% 91.9% 91.4% Random F orest 97.5% 97.4% 99.4% 98.4% Neural Network 94.9% 98.4% 95.0% 96.7% T ABLE IV: Classification performance comparison of dif fer- ent classifiers for 0-15 people in the queue. Accuracy Precision Recall F1 Decision Tree 83.8% 88.9% 90.0% 89.4% AdaBoost 87.3% 93.1% 92.5% 92.8% Random F orest 98.7% 99.4% 100.0% 99.7% Neural Network 97.8% 99.5% 99.2% 99.4% with 3 and 4 persons per square meter respectiv ely , and has a best mean accuracy (98.7%) in the queueing counting. The accuracy , precision, recall and F1 are all abo ve 97% on random forest, demonstrating the extremely satisf actory classification performance. The classification accuracies on 3 persons per square meter are similar to that of 4 persons per square meter on four classifiers, indicating that the hybrid features extraction method is robust and insensitive to the dense en vironments. B. P erformance Comparison with Other F eatures In order to verify the superiority of the proposed hybrid features, three other features are used for comparison, includ- ing the cluster features proposed in [7], the activity features in [10] and the features learnt automatically from LeNet-5 con volutional neural network (CNN) [11]. The cluster features are composed by the detected amplitudes and distances of the corresponding cluster , with the size of 1 × 1280. The acti v- ity features consist "acti vity event" and "acti vity duration", extracted with the size of 1 × 2957. The CNN features are extracted by using the LeNet-5 neural network, which is trained on radar samples for 20 epoches. Features are extracted from the fully connected layer with the size of 1 × 500. The comparison results are conducted with random forest in three dense scenarios, described in Fig. 6. As illustrated, the classification accuracies of proposed hybrid features are obviously better than those of three other features, and ha ve a distinct adv antage of stable performance in three dense scenarios, especially in more complex scene of 4 persons per square meter . Results rev eal that the proposed hybrid features are significantly superior to other features in dense people counting. Fig. 6: Classification accuracies comparison on different features in three dense scenarios. V . C O N C L U S I O N In this letter , a hybrid CTF-DBF feature extraction method for dense people counting with IR-UWB radar is proposed. Features with multiple scales and multiple orientations are extracted from the radar matrix by applying the curvelet transform. Moreo ver , the distance bin is introduced to divide each row of the matrix into sev eral bins along the propagating distance to select features as supplementary information. The dataset in three dense scenarios is constructed, and four classifiers are compared. Counting accuracies are all above 97% with random forest. Moreov er , three other features are compared to verify the superiority of the hybrid features. Comparison results prov e the effecti veness and rob ustness of the proposed method in dense scenarios. In the future work, more radar samples will be collected in more complex scenarios to validate the robustness of the proposed method. R E F E R E N C E S [1] F . Fioranelli, M. Ritchie and H. Griffiths, "Classification of Un- armed/Armed Personnel Using the NetRAD Multistatic Radar for Micro- Doppler and Singular V alue Decomposition Features," in IEEE Geo- science and Remote Sensing Letters , vol. 12, no. 9, pp. 1933-1937, Sept. 2015. [2] Y . Kim, S. Ha and J. Kwon, "Human Detection Using Doppler Radar Based on Physical Characteristics of T argets," in IEEE Geoscience and Remote Sensing Letters , vol. 12, no. 2, pp. 289-293, Feb . 2015. [3] H. Lv et al., "An Adaptive-MSSA-Based Algorithm for Detection of T rapped V ictims Using UWB Radar ," in IEEE Geoscience and Remote Sensing Letters , vol. 12, no. 9, pp. 1808-1812, Sept. 2015. [4] Y . Zhong, Y . Y ang, X. Zhu, E. Dutkiewicz, Z. Zhou and T . Jiang, "Device- Free Sensing for Personnel Detection in a Foliage Environment," in IEEE Geoscience and Remote Sensing Letters , vol. 14, no. 6, pp. 921-925, June 2017. [5] J. W . Choi, S. S. Nam, and S. H. Cho, "Multi-Human Detection Algorithm Based on an Impulse Radio Ultra-W ideband Radar System," in IEEE Access , vol. 4, pp. 10300-10309, 2016. [6] S. Bartoletti, A. Conti and M. Z. Win, "Device-Free Counting via Wide- band Signals," in IEEE Journal on Selected Ar eas in Communications , vol. 35, no. 5, pp. 1163-1174, May 2017. [7] J. W . Choi, D. H. Y im and S. H. Cho, "People Counting Based on an IR-UWB Radar Sensor ," in IEEE Sensors J ournal , v ol. 17, no. 17, pp. 5717-5727, Sept.1, 1 2017. [8] H. Idrees, I. Saleemi, C. Seibert and M. Shah, "Multi-source Multi-scale Counting in Extremely Dense Crowd Images," 2013 IEEE Conference on Computer V ision and P attern Recognition , Portland, OR, 2013, pp. 2547-2554. [9] E. Cianca, M. De Sanctis and S. Di Domenico, "Radios as Sensors," in IEEE Internet of Things Journal , vol. 4, no. 2, pp. 363-373, April 2017. 6 [10] Jin He and A. Arora, "A regression-based radar-mote system for people counting," 2014 IEEE International Conference on P ervasive Computing and Communications (P erCom) , Budapest, 2014. [11] L. N. Smith, E. M. Hand, and T . Doster, "Gradual DropIn of Layers to T rain V ery Deep Neural Networks," in IEEE Conference on Computer V ision and P attern Recognition (CVPR) , 2016. [12] J. Li, L. Liu, Z. Zeng and F . Liu, "Advanced Signal Processing for V ital Sign Extraction With Applications in UWB Radar Detection of T rapped V ictims in Complex Environments," in IEEE Journal of Selected T opics in Applied Earth Observations and Remote Sensing , v ol. 7, no. 3, pp. 783-791, March 2014. [13] Z. Zhang, X. Zhang, H. Y u and X. Pan, "Noise suppression based on a fast discrete curvelet transform," in J ournal of Geophysics and Engineering , 2010, 7(1):105. [14] J. Zheng, Q. Xu, J. Chen and C. Zhang, "The On-Orbit Noncloud- Covered W ater Region Extraction for Ship Detection Based on Relative Spectral Reflectance," in IEEE Geoscience and Remote Sensing Letters , 2018.