A Centralized Processing Framework for Foliage- Penetration Human Tracking in Multistatic Radar

Transcription

1 98 JUN ZHANG, TIAN JIN, YUAN HE, ET AL., A CENTRALIZED PROCESSING FRAMEWORK FOR FOLIAGE-PENETRATION A Centralized Proceing Framework for Foliage- Penetration Human Tracking in Multitatic Radar Jun ZHANG, Tian JIN, Yuan HE, Zhimin ZHOU College of Electronic Science and Engineering, National Univerity of Defene Technology, Deya Street 109, Changha, P. R. China zhangjunnudt@163.com, tianjin@nudt.edu.cn Manucript received November 23, 2015 Abtract. A complete centralized proceing framework i propoed for human tracking uing multitatic radar in the foliage-penetration environment. The configuration of the multitatic radar ytem i decribed. Primary attention i devoted to time of arrival (TOA) etimation and target localization. An improved approach that take the geometrical center a the TOA etimation of the human target i given. The minimum mean quare error paring (MMSEP) approach i introduced for multi-target localization in the multitatic radar ytem. An improved MMSEP algorithm i propoed uing the maximum velocity limitation and the global nearet neighbor criterion, efficiently decreaing the computational cot of MMSEP. The experimental reult verify the effectivene of the centralized proceing framework. Keyword Multitatic radar, foliage penetration, human tracking, TOA etimation, target localization 1. Introduction Human target detection and tracking have great potential in military, afety, ecurity and entertainment application [1]. A number of technologie have been ued for human detection, including computer viion [2], lidar [3], radio frequency identification device (RFID) [4] and radar [1], [5 7]. Computer viion and lidar performance i degraded in duty, foggy and non-line-of-ight environment. RFID i a cooperative enor and can t be ued in noncooperative condition. Radar i a privileged tool to detect human in poor viibility condition (e.g., at night, fog, moke, and in through-wall and foliage-penetration application). Multitatic radar with a low frequency wideband tranmitting ignal ha fine localization preciion, a large covered area, and the ability to penetrate the foliage [5], [8]. It i highly uitable for foliage-penetration urveillance. A multitatic radar ytem uually conit of one tranmitting antenna and everal patially ditributed receiving antenna. Due to cot and ize limitation, antenna of uch a ytem generally have a large beam angle and can only provide limited angle-of-arrival information. By the high range reolution of wideband ignal, time-of-arrival (TOA) i commonly ued for target localization [6], [9]. Multi-target tracking in multitatic radar i a complex tak that require clutter uppreion, target detection, TOA etimation, target localization and target tracking. Human detection i full of challenge due to the low reflectivity of the human body and the evere clutter of the foliage-penetration environment. An appropriate clutter uppreion algorithm hould be ued to remove mot clutter and allow the repone of the moving human to be detected. Target detection and TOA etimation provide the range meaurement for target localization. They hould inure relatively low mi and fale detection rate, and accurate TOA etimation. Target localization i a key challenge in multitatic radar ytem. Epecially in the multiple-target environment, it i neceary to identify, among all available bitatic range meaurement of all receiving channel, which combination of meaurement correpond to true target. Wrong meaurement-to- meaurement aociation lead to the well-known ghot iue [10], [11]. For human target tracking, a decentralized framework ha been propoed in [12]. It ue the human low-time feature for meaurement aociation. The relatively expenive computation cot and the non-real-time tracking limit the application of the decentralized framework [9], [12]. If the ditance between receiving antenna i mall enough, the meaurement aociation method propoed in [6] can be applied. However, thi configuration ha limited localization preciion. In thi paper, we conider ome more general approache for meaurement aociation in multitatic radar ytem. The S-D aignment algorithm i an option, but the aignment problem i NP hard for S 3 (S i the number of receiving channel) [13]. An efficient method propoed in [14] define a lit of potential target baed on a certain metric (a potential target correpond to a combination of the meaurement from each receiving channel). It avoid reolving the optimum aignment problem and identifie the maximum likelihood target by re- DOI: /re APPLICATIONS OF WIRELESS COMMUNICATIONS

2 RADIOENGINEERING, VOL. 25, NO. 1, APRIL peatedly electing the highet potential target on the lit. Thi method need to compute the metric for all the poible combination and ha heavy computational cot. We will offer an improved algorithm in Sec. 3. Target tracking provide the reult of the moving trajectorie of human. It can be implemented by the Kalman filter-baed multitarget tracking (MTT) ytem decribed in [15], [16]. The paper i organized a follow. Firt, the multitatic radar ytem i decribed in Sec. 2. Then, a centralized framework for foliage-penetration human tracking i illutrated in detail in Sec. 3. In Sec. 4, the performance of the propoed approach i evaluated uing the experimental data. Finally, ome concluion are drawn in Sec Multitatic Radar Sytem The diagram of our multitatic radar ytem i preented in Fig. 1. The multitatic radar conit of one tranmitting antenna and three receiving antenna. Antenna are connected to the tranmitter and receiver by cable with length limited to 6 m. The tranmitting ignal i a linear frequency modulation continuou wave with bandwidth 150 MHz at ultra-high frequency band. The radar ha a range reolution of 1 m and i able to penetrate buhe to detect concealed moving target. The receiving ignal i mixed with the tranmitting ignal to reduce the ampling rate in the receiver. Detailed parameter are provided in Tab. 1. Parameter Value Unit Central frequency 800 MHz Bandwidth 150 MHz Pule repetition frequency 1000 Hz Sampling rate 10 MHz Antenna beam width 60 deg. Tab. 1. Operating parameter of multitatic radar ytem. The ampling data of each receiving channel are organized a a 2-D matrix. The column of the matrix repreent the ampling of the received ignal in one pule repetition interval (fat time), and conecutive pule are tored in ucceive column (low time). By performing fat Fourier tranform to the column of each matrix, the range profile can be obtained. To improve the ignal to noie ratio, we divide the low time dimenion into multiple coherent proceing interval (CPI) and add the range profile in each CPI. The data after accumulating proceing are y (t, τ), where t repreent the propagation time, τ repreent the low time whoe interval i a CPI, and the upercript = 1,2,3 repreent the data of channel. 3. Centralized Tracking Framework A decentralized tracking i not capable of real-time tracking [9], we deign a centralized tracking framework for real-time human tracking in multitatic radar. The centralized tracking framework i preented in Fig. 2. Clutter uppreion, target detection and TOA etimation are performed in each receiving channel. The bitatic range meaurement, which can be derived from the TOA value, are fued to etimate the target poition by a multitarget localization algorithm. Finally, MTT i performed to obtain target trajectorie. Each tep i decribed in detail in the following. Fig. 1. Diagram of multitatic radar ytem. 3.1 Clutter Suppreion Generally, the received ignal trength of the human body i weaker than the buh clutter, the antenna coupling and other ambient tatic clutter. Due to the ignal attenuation of buhe and leave, the human echoe become much weaker. The clutter uppreion phae i eential to remove mot clutter and allow the repone of moving human to be detected. There are everal background ubtraction approache for clutter uppreion, uch a accumula- 1 y t, y 2 t, y 3 t, Fig. 2. Centralized tracking framework.

3 100 JUN ZHANG, TIAN JIN, YUAN HE, ET AL., A CENTRALIZED PROCESSING FRAMEWORK FOR FOLIAGE-PENETRATION tive average, moving average, and exponential average [17]. We choe exponential average background ubtraction in our application for it robut performance and low complexity. It implementation can be found in [17], [18]. The data after exponential average background ubtraction proceing are y c (t, τ), where the ubcript c repreent the output of the clutter uppreion phae. 3.2 Target Detection The purpoe of detection i to decide whether target are preent in the examined radar ignal. Contant fale alarm rate (CFAR) detector provide good and robut performance for human detection in wide band radar [7]. Baed on the Neyman-Peron criterion, they have the maximum probability of detection for a given fale alarm rate [19]. Among them, cell averaging CFAR (CA-CFAR) i the mot common detector for moving target detection [9], [20]. However, in multi-target cenario, nearby target will increae the threhold leading to high midetection rate. We apply ordered tatitic CFAR (OS-CFAR) becaue it ha better detection performance in the multi-target environment [7], [21]. The output of the OS-CFAR detector i given a follow: 2 1,, yc t T yd t, (1) 2 0, yc t, T where T i the adaptive threhold provided by OS-CFAR, the ubcript d repreent the output of target detection phae. After OS-CFAR, one range profile correpond to a vector with binary value. 3.3 TOA Etimation A human target i a typical exted target. With the range reolution of 1 m, the fat time repone of a normalized peron cover about 4 m in our multitatic radar ytem. To improve the accuracy of TOA etimation, we et the bitatic range ampling interval to 0.2 m. Therefore, one peron i uually repreented by everal TOA value. To implify the target localization, each detected peron hould correpond to only one TOA. A method that take the leading edge a the TOA etimation of the human target ha been propoed in [6]. Due to the pecific motion characteritic of people, leading edge may indicate the arm and leg, which move ahead of toro. Additionally, the range reolution of 1 m i not fine enough to ue the leading edge TOA etimation method. An improved approach i to elect the geometrical center (the center of the exted TOA value of the human) a the TOA etimation. If the length of the filter doe not match the human extenion or if the ditance of the target in the range profile i cloe, the average filter method may generate unwanted TOA etimation [9]. Here, we propoe a robut geometrical center TOA etimation method. The geometrical center TOA etimation method conit of three tep. The firt two tep are the ame a the leading edge TOA etimation method in [6]. A the parameter of our ytem are different from [6], we give whole tep a follow: 1. Accumulation of detector output with interval length correponding to the exted ize of the target. interval: Summation of, y t along the propagation time d Wp 1 e i d in n0 y ( t, ) y ( t, ) (2) where W p repreent the exted ize of the target, and the ubcript e repreent the TOA etimation phae. A the bitatic range ampling interval i et to 0.2 m, the exted ize 4 m of a normal peron correpond to the interval length W p = Generation of continuou TOA interval. Comparing y ( t, ) with a elected threhold T lim : e i 1, ye ( t, ) Tlim y e (, t ) (3) 0, ye ( t, ) Tlim where T lim determine the minimal number of reflection that confirm a target. T lim i et to 8 in our ytem. After the above two tep, the majority of the fale alarm and midetection can be mitigated. 3. Subtitution of exted TOA value by ingle TOA etimation. In thi tep, each continuou TOA interval i ubtituted by, at mot, two TOA etimation. The peudocode procedure of geometrical center TOA etimation i given in Tab. 2. The variable N t i the total number of ample of y e (, t ) in the range profile. TOA m(τ) repreent the geometrical center TOA etimation of the mth target at low time τ in channel. The operator x give the integer nearet to x. If two target are cloe enough in range, their exted TOA value may overlap and caue continuou TOA interval. The variable tnum conider the above cae, and if tnum 2, it i predicted that there are two target in the continuou TOA interval. When the loop, the number of target and their correponding TOA etimation can be etimated. The performance of the TOA etimation method will be verified by the TOA etimation reult and human tracking reult in Sec. 4. By multiplying the TOA etimation with the electromagnetic wave propagation velocity c, bitatic range meaurement are given:, 1,2, r ctoa m m (4) m m where m i the etimated target number in channel. Target localization can be performed by the Taylor erie

4 RADIOENGINEERING, VOL. 25, NO. 1, APRIL algorithm [9], [22] with the elected bitatic range meaurement correponding to the target in each channel. Then, the quetion i raied of which combination of meaurement in the receiving channel correpond to true target. Wrong meaurement aociation will lead to the wellknown ghot iue [10], [11]. S 0, E 0, m0 for i 1 Nt2 if [ y ( t, ) 1] AND [ y ( t, ) 1] ele e i e i1 AND [ y e ( ti2, ) 1] if ( S 0) S i, E i ele E E1 if S ~ 0 tnum ( E S 2 Tlim Wp)/( Wp Tlim ) if ( tnum 2) TOA t TOA mm2 ele TOA m1 SWp /2Tlim m2 t E3 Wp / 2Tlim mm1 S 0, E 0 t m1 ( SEWp )/2 Tab. 2. Peudocode procedure of geometrical center TOA etimation. 3.4 Target Localization For multi-target localization in a multitatic radar ytem, one ha to identify, among all available bitatic range meaurement of all receiving channel, which combination of meaurement correpond to true target. An efficient method ha been propoed for range-to-range aociation in multi-target multi-enor environment [14]. We introduce thi method for our meaurement aociation problem and propoe an improved algorithm. Aume that the number of output range meaurement in each channel are m 1, m 2, and m 3, repectively. We will conider three meaurement from three receiving channel, making the total number of poible combination (potential target) in the initial lit m 1 m 2 m 3. One can calculate a preumed poition of the target uing the elected combination by a certain localization algorithm. Then, a metric can be calculated that i baed on all the range meaurement upon the elected combination. Here, we choe mean quare error (MSE) a the metric. For intance, a potential target at (x, y) can be etimated upon a combination {r 1 i1, r 2 i2, r 3 i3} (r i repreent the i th meaurement in receiving channel ). The MSE metric i defined a follow: MSE 3 1 ( r r ) where r ( xxt) ( y yt) ( xx) ( y y) i the etimated bitatic range, (x t, y t ) i the poition of the tranmitting antenna, and (x, y ) ( = 1,2,3,) i the poition of the receiving antenna. Firt, we can et a maximum permitted MSE and delete the combination with MSE larger than the permitted MSE. Then, the lit of potential target i ordered by the MSE metric, with the minimum MSE at the top and the maximum MSE at the bottom. Thi approach aume that the potential target highet on the lit ha the maximum likelihood of being the true target. Select the potential target highet on the lit and eliminate all other lower ordered potential target that are baed on any of the meaurement upon which the elected target wa baed. The lit can thu be pared down. The operation i repeated for the next highet potential target remaining on the lit until the number of elected target reache the preet target number or the lit i empty. A the method elect the target with minimum MSE and pare down the lit in each proce loop, we call it minimum MSE paring (MMSEP) algorithm. The MMSEP algorithm can be ummarized by the following tep: 3 i 2 (5) 1. Lit all the combination, etimate the poition of potential target, and calculate the MSE of each combination. Delete the combination with MSE larger than the permitted MSE. 2. Order the lit of potential target by their MSE, with the minimum MSE at the top and the maximum MSE at the bottom. 3. Select the potential target highet on the lit and eliminate all other lower ordered potential target baed on any of the meaurement upon which the elected target wa baed. Add the elected potential target to a lit of true target. 4. Repeat tep 3 until the number of elected target reache the preet target number or the lit i empty. The main computational cot lay in the firt tep, which mut compute the location and the MSE of all poible combination. In thi paper, we propoe the following method to reduce the computational cot in the original algorithm.

5 102 JUN ZHANG, TIAN JIN, YUAN HE, ET AL., A CENTRALIZED PROCESSING FRAMEWORK FOR FOLIAGE-PENETRATION Auming that the Carteian location at low time k 1 i: Pk-1 x j, yj, j 1,2, mk (6) 1 where m k - 1 i the etimated number of the target at low time k 1. Then, the bitatic range of target j in each receiving channel can be calculated: r ( x x ) ( y y ) ( x x ) ( y y ). (7) j j t j t j j The bitatic range meaurement in each receiving channel at low time k are: Rk rj, jk 1,2, m (8) k where m k i the number of meaurement in receiving channel at low time k. We define the following inequality: min r r 2V T r (9) jk jk j where V max i the maximum moving velocity of the human (normally V max < 10 m/), r i a preet range error, and T i the time interval between low time k 1 and k (i.e., CPI). A the moving ditance in T i limited by 2V max T (2V max i the maximum bitatic velocity [23]), the meaurement atifying inequality (9) have the highet likelihood to be aociated with target j. The poition of j at low time k can be calculated baed on the aociated meaurement. Delete the aociated meaurement from R k and repeat aociation for other target in P k 1 until j = m k 1 or until at leat one et among R k, = 1, 2, 3 i empty. After above proceing, if all et R k, = 1, 2, 3 are not empty, then perform MMSEP algorithm upon remaining meaurement. A the global nearet neighbor (GNN) criterion i ued in inequality (9), we call the propoed procedure a GNN- MMSEP algorithm. The GNN-MMSEP algorithm i ummarized a follow: 1. Perform the MMSEP algorithm for target localization at low time k = 1. The poition of the target are tored in et P 1,k = k Initialize j = Calculate r j, and min rj k rj. If there are three jk bitatic range meaurement in three receiving channel at low time k that atify the inequality (9), calculate the poition upon the meaurement, add the poition to et P k, and delete the meaurement in R k. 4. Perform tep 5 if j = m k 1 or at leat one et among R k, = 1, 2, 3 i empty; otherwie j = j + 1 and return to tep If all et among R k, = 1, 2, 3 are not empty, perform MMSEP algorithm for target localization upon remaining meaurement and add the poition to P k. k = k + 1. Return to tep 2. max k The GNN-MMSEP algorithm perform MMSEP for target localization at the firt low time interval. At ucceive low time interval, it ue GNN criterion for target aociation to decreae the computational cot of target localization for all poible combination in MMSEP. The 5th tep maintain the target that are not aociated with the target at the previou low time interval to be located. There are two factor that can affect the target localization performance of the GNN-MMSEP algorithm. 1. If there are intenive clutter meaurement in the receiving channel, inequality (9) may aociate clutter meaurement with the target. A the TOA etimation method above can reject clutter, the influence of clutter can be omitted. Additionally, wrong aociation will generate fale poition at current low time interval and the fale poition ha le poibility to be aociated with the target at ucceive low time interval. The true target poition will be relocated in tep 5. The algorithm itelf i robut to pare clutter environment. 2. If target are cloe in range, it alo may caue wrong aociation. Our algorithm i alo robut in thi cae, a the poition calculated by wrong combination at current low time interval i le likely to be aociated with the target at ucceive low time interval. The true target poition will be relocated in tep 5. The experimental reult in Sec. 4 demontrate the performance of the GNN-MMSEP algorithm. 3.5 Target Tracking The location of the target calculated by the range meaurement normally have certain random error. Generally, target tracking provide more accurate and moothing trajectorie. For MTT, all the meaurement-to-track aociation, track filtering, track initialization and track maintenance tep hould be conidered. We choe a Kalman filter baed MTT framework, a it ha been demontrated robut and efficient for human tracking [15]. Detailed decription of the implementation of MTT ytem can be found in [15], [16]. 4. Experimental Reult 4.1 Experimental Scenario The multitatic radar i localized behind a thick and impenetrable buh. Two people move on the other ide of the buh. The buh i approximately 2 m thick, 1.8 m high and 9 m wide. The poition of antenna are et a follow: TX (1,0), RX1 (-4,2), RX2 (-1,1) and RX3 (4,2), where TX/RX indicate the tranmitting antenna and receiving antenna, repectively. The antenna were et to inure that they are not in line-of-ight of the moving human. The experimental cenario i illutrated in Fig. 3. The data ued

6 RADIOENGINEERING, VOL. 25, NO. 1, APRIL Fig. 3. Experimental cenario. (a) The foliage-penetration cenario. (b) Moving human. (c) Multitatic radar etup. in thi paper were recorded in a cloudy day with gentle breeze in September. The relative humidity of the air wa about 60%. The buh had buhy leave. (a) 4.2 Reult The low time-range image of the receiving channel one i preented in Fig. 4. Figure 4(a) i the image before clutter uppreion. The human trajectorie are ubmerged in the clutter and difficult to be detected. By exponential average background ubtraction clutter uppreion method, mot clutter i removed and human trajectorie are clearly revealed in Fig. 4(b). Note that the performance of clutter uppreion in the foliage-penetration environment relie heavily on the weather. Further tudy for clutter uppreion will be taken under different weather. (b) (a) Fig. 4. Time-range image of receiving channel one. (a) Before clutter uppreion. (b) After clutter uppreion. (a) Fig. 5. TOA etimation reult of receiving channel one. (a) Leading edge method. (b) Geometrical center method (yellow circle are the TOA etimation of the people). (b) (b) (c) Fig. 6. Tracking reult. (a) MMSEP localization with leading edge TOA etimation. (b) MMSEP localization with geometrical center TOA etimation. (c) GNN-MMSEP localization with geometrical center TOA etimation (red line are the ground truth, green circle are the localization reult, and blue line are the tracking reult). The OS-CFAR detection and the TOA etimation procedure are performed on the range profile after clutter uppreion. Figure 5(a) i the reult of the leading edge method propoed in [6]. It take the leading edge of the OS-CFAR output a the TOA etimation of the target,

7 104 JUN ZHANG, TIAN JIN, YUAN HE, ET AL., A CENTRALIZED PROCESSING FRAMEWORK FOR FOLIAGE-PENETRATION which may produce error a decribed in Sec. 3.3, and it may mi the meaurement of the further target when target are cloe to each other in the range direction. Figure 5(b) i the reult of our geometrical center TOA etimation method. It can provide a relatively accurate TOA etimation with fewer miing meaurement. Figure 6(a) and 6(b) are the tracking reult of MMSEP localization uing the leading edge and geometrical center TOA etimation method. The red line repreent the ground truth of the moving people (Peron A walk from P1 to P2; Peron B walk from P3 to P4), the green circle are the localization reult of the MMSEP algorithm, and the blue line are the tracking reult of the MTT ytem. The dicontinuitie of the trajectorie in Fig. 5 are due to the loe of range meaurement in the cae of target trajectory croing. A the leading edge TOA etimation loe more meaurement, the dicontinuitie are more eriou than in the geometrical center method. The tracking error of the geometrical center TOA etimation are maller hown in Fig. 6(b), which further demontrate the uperiority of the geometrical center method. We have alo performed the GNN-MMSEP algorithm upon the meaurement of the geometrical center TOA etimation for target localization. Figure 6(c) how the localization and tracking reult of GNN-MMSEP. When target cro in the range profile, thi algorithm may aociate wrong meaurement, a decribed in Sec Wrong meaurement combination will lead to wrong localization. Therefore, the dicontinuitie of the trajectorie in Fig. 6(c) are lightly wore than Fig. 6(b). However, thi mall gap can be tolerated in human tracking application. Although comparable localization error wa preented in Fig. 6(b) and Fig. 6(c), GNN-MMSEP approach indeed demontrated a lower computational cot in comparion with the original MMSEP approach. In our experimental cenario, the total proceing time of 120 low time interval by MMSEP i However, the GNN-MMSEP algorithm took only (CPU: Intel i5-3470, MATLAB verion: R2012b). By uing the GNN-MMSEP algorithm, we improve the computational efficiency of target localization by more than 20 time. To evaluate the performance of the tracking framework, a mean tracking error metric i defined a: Ee E x x 2 y y 2 (10) where (x(τ), y(τ)) i the peron ground truth at low time τ (we aume that the human are walking at a contant velocity), and (x (τ), y (τ)) i the peron predicted poition provided by the tracking reult. Table 3 how the tracking preciion of different TOA etimation and target localization method. The tracking framework uing leading edge TOA etimation ha the wort performance. The mean tracking error i more than 1 m. By geometrical center TOA etimation, the tracking framework with the MMSEP and GNN-MMSEP location algorithm nearly have the ame tracking preciion. TOA etimation Mean tracking error Location method method Peron A Peron B Leading edge MMSEP 1.02 m 1.26 m Geometrical center MMSEP 0.28 m 0.2 m Geometrical center GNN-MMSEP 0.3 m 0.19 m Tab. 3. Mean tracking error between tracking reult and the ground truth. 5. Concluion In thi paper, we have propoed a complete centralized proceing framework for foliage-penetration human tracking in multitatic radar. A geometrical center method i propoed for the more accurate TOA etimation of human target. The MMSEP algorithm i introduced for more general human target localization in a multitatic radar ytem. The propoed GNN-MMSEP method ha hown better performance in term of computational efficiency in comparion with the original MMSEP algorithm. The experimental reult how the validity of the propoed centralized tracking framework for foliage-penetration human tracking. The introduced centralized tracking framework alo ha great potential for other application, uch a through-wall target tracking, air vehicle tracking, and hip tracking in a harbor. Acknowledgment Thi work wa upported in part by the National Natural Science Foundation of China under Grant and the reearch project of National Univerity of Defene Technology under Grant CJ Reference [1] CHANG, S.H., MITSUMOTO, N., BURDICK, J. W. An algorithm for UWB radar-baed human detection. In IEEE Radar Conference. Paadena (CA, USA), 2009, p DOI: /RADAR [2] TAJ, M., CAVALLARO, A. Multi-view multi-object detection and tracking. Computer Viion, Springer Berlin Heidelberg, 2010, vol. 285, p DOI: / _10 [3] HASHIMOTO, M., TSUJI, A., NISHIO, A., et al. Laer-baed tracking of group of people with udden change in motion. In IEEE International Conference on Indutrial Technology. Seville (Spain), 2015, p DOI: /ICIT [4] SÄRKKÄ, S., VIIKARI, V. V., HUUSKO, M., JAAKKOLA, K. Phae-baed UHF RFID tracking with nonlinear Kalman filtering and moothing. IEEE Senor Journal, 2012, vol. 12, p DOI: /JSEN [5] LEI, P. Z., HUANG, X. T. Robut detection of moving human target in foliage-penetration environment baed on Hough tranform. Radioengineering, 2014, vol. 23, no. 1, p ISSN: [6] ROVŇÁKOVÁ, J., KOCUR, D. TOA etimation and data aociation for through-wall tracking of moving target. Euraip

8 RADIOENGINEERING, VOL. 25, NO. 1, APRIL Journal on Wirele Communication and Networking, 2010, p DOI: /2010/ [7] URDZIK, D., KOCUR, D. CFAR detector for through wall tracking of moving target by M-equence UWB radar. In 20th International Conference Radioelektronika. Brno (Czech Rep.), 2010, p DOI: /RADIOELEK [8] ZETIK, R., JOVANOSKA, S., THOMA, R. Simple method for localization of multiple tag-free target uing UWB enor network. In IEEE International Conference on Ultra-Wideband. Bologna (Italy), 2011, p DOI: /ICUWB [9] HE, Y. Human target tracking in multitatic ultra-wideband radar. Ph.D. Diertation. Delft Univerity of Technology, ISBN: DOI: /uuid:3fd16a a-bba3-56fcf49f4b37 [10] BOSSE, J., KRASNOV, O., YAROVOY, A. A imple direct target localization and deghoting algorithm in active radar network. In 2014 International Radar Conference. Lille (France), 2014, p DOI: /RADAR [11] DAUN, M. Deghoting in paive air urveillance ytem. In IEEE 11 th International Radar Sympoium. Vilniu (Lithuania), 2010, p ISSN: [12] HE, Y., AUBRY, P., LE CHEVALIER, F., YAROVOY, A. G. Decentralized tracking for human target in multitatic ultrawideband radar. IET Radar, Sonar and Navigation, 2014, vol. 8, no. 9, p DOI: /iet-rn [13] PATTIPATI, K. R., SOMNATH, D., BAR-SHALOM, Y., WASHBURN, JR. R. B. A new relaxation algorithm and paive enor data aociation. IEEE Tranaction on Automatic Control, 1992, vol. 37, no. 2, p DOI: / [14] PARADIE, M. J., LABITT, B. D. Method and apparatu for identifying complex object baed on range reading from multiple enor. U.S. Patent 6,664,918, [15] ROVŇÁKOVÁ, J. Multiple target tracking ytem for through wall application. In Proceeding of the Scientific Conference of Young Reearcher. Košice (Slovak Republic), 2010, p [16] KHAN, J., NIAR, S., RIVENQ-MENHAJ, A., HILLALI, Y. E. Multiple target tracking ytem deign for driver aitance application. In Proceeding of the Deign and Architecture for Signal and Image Proceing. Bruel (Belgium), November [17] PICCARDI, M. Background ubtraction technique: A review. In Proc. IEEE International Conference on Sytem, Man and Cybernetic. 2004, vol. 4, p DOI: / ICSMC [18] ZETIK, R., CRABBE, S., KRAJNAK, J., PEYERL, P., SACHS, J., THOMA, R. Detection and localization of peron behind obtacle uing M-equence through-the-wall radar. Proc. SPIE 6201 Senor, and Command, Control, Communication, and Intelligence (C3I) Technologie for Homeland Security and Homeland Defene V. Orlando (FL, USA), DOI: / [19] KAY, S. M. Fundamental of Statitical Signal Proceing, Vol. II: Detection Theory. Signal Proceing, Upper Saddle River, NJ: Prentice Hall, ISBN: [20] RICHARDS, M. A., SCHEER, J. A., HOLM, W. A. Principle of Modern Radar: Baic Principle. Raleigh (NC): SciTech Publihing, ISBN: [21] ROHLING, H. Radar CFAR threholding in clutter and multiple target ituation. IEEE Tranaction on Aeropace and Electronic Sytem, 1983, vol. 19, no. 4, p DOI: /TAES [22] SVECOVA, M., KOCUR, D., ZETIK, R. Object localization uing roundtrip propagation time meaurement. In 18th International Radio electronic Conference.Prague (Czech Rep.), 2008, p DOI: / RADIOELEK [23] WILLIS, N. J. Bitatic Radar. Raleigh (NC): SciTech Publihing, ISBN: About the Author... Jun ZHANG wa born in Yancheng, Jiangu, China. He received the B.S. degree in Information and Telecommunication Engineering from Xidian Univerity, Xi an, China, in 2010, and the M.S. degree in Electronic Information Science and Technology from the National Univerity of Defene Technology, Changha, China, in He i currently working toward the Ph.D. degree in Information and Telecommunication Engineering at the National Univerity of Defene Technology. Hi current reearch interet include multitatic radar ytem deign, target detection and target tracking. Tian JIN received the B.S., M.S., and Ph.D. degree from the National Univerity of Defene Technology, Changha, China, in 2002, 2003, and 2007, repectively, all in Information and Telecommunication Engineering. Hi Ph.D. diertation wa awarded a the National Excellent Doctoral Diertation of China in He i currently a Profeor in the National Univerity of Defene Technology. Hi reearch interet include radar imaging, automatic target detection, and machine learning. Yuan HE wa born in Guangyuan, Sichuan, China. He received hi B.S. and M.S. degree in Information and Telecommunication Engineering from the National Univerity of Defene Technology, Changha, China in 2007 and 2010, repectively. He received hi Ph.D. degree in Telecommunication Engineering from Delft Univerity of Technology, the Netherland in He i currently a reearcher with the National Univerity of Defene Technology. Hi reearch interet include radar ytem deign, ignal proceing, and large-cale data cience. Zhimin ZHOU received the B.S., M.Sc., and Ph.D. degree in Information and Communication Engineering from the National Univerity of Defene Technology, Changha, China, in 1982, 1989, and 2002, repectively. He i currently a Profeor with the National Univerity of Defene Technology. Hi interet include SAR ytem and realtime ignal proceing.