Radar Shadow and Superresolution Features for Automatic Recognition of MSTAR targets

Transcription

1 Radar Shado and Superresolution eatures for utoatic Recognition of STR targets Jingjing Cui, Jon Gudnason, ike Brookes. Iperial College London Key Words: Hidden arkov odel, Target Recognition, High Range Resolution, Synthetic perture Radar, ultiple Signal Classification, eature Extraction, oving and Stationary Target cquisition and Recognition (STR) BSTRCT utoatic target recognition fro high range resolution radar profiles reains an iportant and challenging proble. In this paper, e present a novel feature set for this task that cobines a noise-robust superresolution characterisation of the target scattering centres derived using the USIC algorith ith a representation of the target's radar shado shape. To obtain the shado shape features, three alternative spectral estiation ethods are investigated. Using a Hidden arkov odel to represent aspect dependence, e deonstrate that the inclusion of the shado features results in a significant iproveent in recognition perforance. Using aziuth apertures of 3 and 6 in a -target classification task fro the STR database, e obtain overall classification error rates of.3% and.2% respectively. These results are significantly better than those obtained by other published ethods on the sae database.. INTRODUCTION The autoatic detection and classification of targets fro their radar signatures is an iportant and difficult proble that has attracted considerable research effort. lgoriths for target recognition fro high range resolution (HRR) radar signals generally use as their priary input either a synthetic aperture radar (SR) iage or else a sequence of one or ore one-diensional HRR range profiles. The iage-based approaches generally have higher perforance but are less robust to target otion because of their long data acquisition tie. Soe iage-based algoriths use the pixel values of the iage as their recognition features [, 2, 3] hile others first transfor the iage to another doain [4, 5]. n alternative approach for targets that are large copared ith the radar avelength is to odel the radar return as eanating fro a discrete set of orientation-dependent points knon as scattering centres [6]. In this approach, the SR iage is processed to generate an explicit list of scattering centre positions and associated radar cross sections on hich the recognition features are based [7, 8]. In the sae ay, systes that act on the HRR range profiles can either use the ra [9] or transfored [] profile values as their features or else can process the profiles to estiate the scattering centre locations and cross sections []. Both SR iages and HRR profiles can exhibit large variations for sall changes in target orientation. Target recognition systes ust account for this aspect-dependency by using a rotation invariant transfor [4] or by having ultiple, orientation-dependent, target representations hich ay conveniently be ebedded in a Hidden arkov odel (H) [5,, 2]. In this paper, e present a novel feature set for autoatic target recognition fro a sequence of radar range profiles. Our feature set uses a noise-robust super-resolution technique for identifying scattering centre locations and cobines this inforation ith additional features that characterise the shape of the radar shado. ig. (b) shos a SR iage of a T72 tank taken fro the STR [3, 4] dataset. This iage ay be divided into three regions having significantly different characteristics: (a) the target itself, (b) the target shado and (c) a clutter region surrounding the target. s can be seen in this exaple, the shape of the shado region gives potentially useful inforation about the vertical profile of a target hen it is sited on level ground. This inforation is not available fro the direct target returns hich are insensitive to vertical displaceent. The shado inforation has been used by others to iprove target detection [5] but is not generally used explicitly in target recognition. We shoed in [6] that the shado inforation can be effective in radar target recognition and e present further iproveents in perforance in the present paper HRR Profiles spect ngle [deg] ig.. (a) HRR profiles (b) SR iage of T72 tank SR Iage cross In Section 2 of this paper, e describe our proposed feature set in detail and in Section 3, e describe the Hidden arkov odel that e use to represent the aspect dependency of the radar returns. In Section 4, e evaluate the perforance of our target recognition syste using observation data fro the STR database [4] and copare its perforance ith that of other systes fro the literature that use the sae database. inally, e suarise our results in Section 5.

2 2. RECOGNITION ETURE SET The features that e use for target recognition are derived fro the sequence of coplex-valued HRR profiles, x ( n,, obtained by applying a discrete ourier transfor (DT) to the indoed phase history radar returns. Here n is the profile index and k is the range-bin index covering the region of interest. ig. (a) shos a typical plot of x ( n, and ig. (b) shos the SR iage that results fro indoing x ( n, and taking the DT ith respect to n. Visible in the iage are the target itself, ith signal levels ell above the clutter noise level and also a ell defined shado region ith very lo signal levels. or each value of the profile index n, e obtain to feature vectors: u (n) characterizes the positions and intensities of the scattering centres ithin the target hile (n) characterises the shape of the shado area. Both these feature vectors are derived fro 2 P + consecutive profiles centred on profile n. We therefore define the data atrix x n ( = x( n +, here p { P,..., P}. The procedures to obtain u (n) and (n) are described belo. 2. Target features The derivation of u (n) is illustrated in ig. 2 and the processing steps are described belo here, for clarity, e oit the profile index, n. x( range ask for target USIC in cross range a,k,k aziuth ask y(l, log & dct select lo frequency coponents ig.2. Procedure to calculate target features fro HRR profiles Each scattering centre in range bin k gives rise to a coplex exponential ter in x ( n, and the first step in obtaining u is to identify these ters. We do this by applying the USIC algorith [7] hich uses the data odel: x( = α, e + ε( k here α and ω are the coplex aplitudes and frequencies of the scattering centre ters and ε ( is assued to be hite noise. The reasons for using the USIC algorith are that it is resistant to noise, does not require indoing of x ( n, and is able to estiate ω ith high resolution independently of P. Within range bin k, each of the coponents in () corresponds to a scattering centre hose cross-range displaceent is proportional toω. The axiu nuber of scattering centres,, could be chosen adaptively for each range bin but e have, in the experients belo, fixed it for each range bin according to the nuber of HRR profiles used. fter discarding any scattering centres hose cross-range displaceent lies outside the target ask, e convert the jω p u () continuous displaceents, ω, to discrete values. We create a continuous signal containing an ipulse for each scattering centre hich e then lo-pass filter and saple to give: y( l, = α h( l βω, k ) (2) here h( l) = (.33πl ) sin(.33πl ) is the lo-pass filter response and β =.5λ(2π φ r) is a constant ith λ, φ and r, the avelength, aziuth increent and cross range resolution, respectively. inally, the iage inforation is copressed by taking the 2-diensional discrete cosine transfor (DCT) of log y( l, and e for a 54-eleent feature vector u, retaining only the coefficients in the lo frequency triangle of size -by- and excluding the DC ter. ig. 3(a) shos an exaple of y ( l, corresponding to the central portion of the iage shon in ig. (b) and ig. 3(b) shos the reconstructed logarithic iage using only the coefficients for u Target Iage using USIC cross 2 Reconstruced Target Iage using USIC cross ig. 3 (a) The iage y( l, of the T72 tank (b) The reconstructed logarithic iage using u 2.2 Shado features The shado area of the SR iage consists of deep valleys corresponding to the shado and broad peaks representing the clutter. Unlike the target area of the SR iage, the shado region contains no sharp peaks fro scattering centres and e ish to characterise the shape of its boundary. The procedure for obtaining the shado feature vector,, is shon in ig. 7. olloing a range ask, e perfor spectral estiation in the cross-range direction using one of three ethods described belo. We then identify the shado region using an adaptive threshold and, as ith the target features, copress the inforation using a DCT. x( range ask for shado spectral estiation I(l, threshold b(l, dct select lo frequency coponents ig.7 Procedure to calculate shado features fro HRR profiles using one of the three spectral estiation ethods 2.2. ourier Transfor spectral estiation The first ethod of spectral estiation is to apply a Haing indo to x ( and then to take the DT in the p direction. The shape of the spectru is thus estiated for each range bin k and the iage I( l, is fored in the sae

3 ay as a SR iage. draback of this ethod is that the cross-range resolution is liited by the indoing operation utoregressive spectral estiation better spectral estiate is achieved by deriving the paraeters of a odel of the cross range profile. irst e apply the autocorrelation ethod of autoregressive (R) spectral estiation to x( using the odel: x( = a ( x( p, + u( (3) here a ( are coplex R coefficients and u( is a hite input driving sequence of zero ean and unit variance. We then for the iage I ( l, by converting the R coefficients k ) to the spectru in the p direction: a ( I( l, = a ( e j2πl The R odel is an all-pole odel and the odel order,, is fixed at 25 in the experients belo. The odel order is high because the peaks in the resulting spectru ill odel the clutter around the shado even though the recognition is not done on these peaks but on the shape of the shado oving average spectral estiation We can also estiate the spectru of x( by applying a oving verage () odel: (4) x( = b ( u( p, (5) here b ( are coplex coefficients and u( is the input driving sequence. The odel is an all-zero odel hich characterises the shado. The odel order, in the experients belo, fixed at 2, in order to locate edge of shado. We for the iage I ( l, by converting the coefficients k ) to the spectru in the p direction: b ( I( l, = b ( e j2πl To estiate the coefficients, e first convert the process to a high-order R process [8] and then use standard R estiation procedures to obtain the coefficients. The ethod uses the folloing odel: N x( = u( an ( x( p n, ( N >> ) n= (6) (5) for p = v( = a p ( for < p N + (6) for p > N + v( = u ( b ( v( p, (7) In our ipleentation e used N = Binary Iage oration fter obtaining the iage I ( l, fro the shado ask, e threshold it to give a binary-valued shado iage b ( l, shon in ig. 5 for each of the spectral estiation ethods. The pixel values representing the clutter region in I ( l, have a large spread hereas the shado pixels are concentrated in loer values. This allos us to deterine the threshold fro the histogra by choosing it to be above the highest value of the bins containing the ost pixels. The shado iages in ig. 5 are obtained fro the data shon in ig. using (a) the ourier transfor, (b) R spectral estiation and (c) spectral estiation. The figures sho that the shado is featured proinently as a large black area hile the clutter is shon as ixed points of black and hite. s ith the target features, e copress the shado iage by taking a 2-diensional DCT and retaining 54 lo frequency coefficients to for the shado feature vectors denoted by, and respectively for the three ethods. The reconstructed iages using, and are shon in ig. 6. The shapes of the shados are all ell retained except that the ethod introduces soe horizontal steaks. Shado Iage using T cross Shado Iage using R cross Shado Iage using cross ig. 5 The iage b ( l, using (a) ourier transfor, (b) R spectral estiation, (c) spectral estiation. Reconstruced Shado Iage using T cross Reconstruced Shado Iage using R cross ig. 6 Reconstructed iages fro (a), (b) Reconstruced Shado Iage using cross, (c) The nuber, 2 P +, of HRR profiles used to for the feature vector has a direct effect on the cross-range resolution of and but not on that of. Hoever although R spectral estiation can achieve superresolution, its error criterion concentrates on the high energy features in the clutter region and so does not find the shado edges accurately. In contrast the approach represents the shado region ell even using only a 2 nd order odel. We copare the classification perforances using the three kinds of features in the Section 4.

4 soothed log likelihood soothed log likelihood 3. ZIUTH HIDDEN RKOV ODEL HRR profiles exhibit significant variability ith target orientation. We odel this for each target using an H containing S states hich correspond to different target aspects. Within an observation sequence, consecutive HRR profiles correspond either to the sae or to adjacent states. Thus the only alloable state transitions are fro a state to itself or to the adjacent state in the direction of sensor otion. We initialise the states to correspond to equal aspect increents of S and for each state e train a Gaussian ixture odel (G) [9] using all available training data fro the corresponding range of aspects of a particular target. The transition probability beteen adjacent states is initialised to be S φ / here φ is the aziuth increent beteen successive feature vectors. Using these initial values, e then re-estiate the G paraeters and the H transition probabilities using ebedded Bau-Welch training []. The aziuth interval represented by a state can change during re-estiation as is illustrated in ig. 8(a). This figure shos the log likelihood of test feature vectors as a function of aziuth angle for each of three consecutive odel states. The three odel states ere initially trained ith data fro consecutive 6 aziuth intervals in the region of 6. We see fro ig. 8(a) that the log likelihood does indeed peak at a target orientation of 6 and that there is a second, saller, State Decoposition peak at around S= S= 25 due to target - S=2 syetry. n - enlarged vie of the priary peak is - shon in ig. 8(b) -4 hich shos that -5 each state has retained clearly aspect angle [degree] 35 4 defined boundaries. State Decoposition State is hoever S= S= uch broader than -2 S=2 its initial idth of 6 and no covers -4 about, state -6 covers a relatively narroer angular -8 extent about 2 and - state 2 becoes aspect angle [degree] broader again. ig. 8 State decoposition (a) fro 25 to to 4. EXPERIENTL RESULTS (b) fro 4. STR database The experiental evaluations ake use of the oving and Stationary Target cquisition and Recognition (STR) database collected by the Sandia National Laboratory using an X-band SR sensor in. 3 resolution spotlight ode [4, 3]. The database contains coplex valued SR iage chips of confusable targets and their variants. or each target, the iages cover a full aziuth range at depression angles of 5 and7 for test and training data respectively. The SR iages have a resolution of r =. 3 in both the range (horizontal) and cross range (vertical) directions. or our experients, the SR iage chips ere converted into a sequence of HRR profiles by taking an inverse ourier transfor, reoving the zero padding and finally undoing the Taylor indo in the P direction. The steps of this procedure are suarised in ig. 9. Each iage chip covers an aziuth interval of approxiately 3 [9] ith successive HRR profiles separated by an angular increent of φ =.3. ig.9. Procedure to convert STR iages to HRR profiles. The operations are in cross-range. 4.2 Experiental procedure or our experients e used an H containing S = 6 states each corresponding to an initial aziuth interval of 6. Within each state, feature vector distributions are represented by a diagonal-covariance Gaussian. We used a 9 9 target ask and a 8 9 shado ask. total of iage chips containing 7 depression angle data as used to train a separate H for each of the targets. or testing, e used a total of 5 5 depression angle iages ithout any copensation for the slight isatch in depression angle. We evaluated three alternative feature sets: the target features u, the shado features and the concatenation of the to u +.The first to sets contain 54 eleents hile the last has 8. The feature vectors ere fored using P = 25 corresponding to an aperture of.5.the axiu nuber of scattering centres,, is fixed at for each range bin. We fored a sequence of HRR profiles covering 3 aziuth aperture fro one iage chip and a sequence covering 6 fro to adjacent iage chips. We then deterined the odel ith the highest likelihood. odel training and recognition ere perfored using the HTK recognition softare []. We are interested in knoing hat ethod to use for extracting the shado features. On their on, the shado features do not perfor ell, but in conjunction ith the target features, they can iprove the recognition results significantly. Table shos a closed-set identification results using the target features u, and three proposed shado features, and used both on their on and concatenated ith u to for a 8 eleent feature vector. The table presents the percentage isclassification rates for each

5 individual target, the overall test set isclassification rate (CR) and the standard deviation of the test set isclassification rate (St.Dev) using test sequences hich each cover a 3 aziuth aperture. We can see that if used on their on, the R derived shado features perfor best ith 7.3 % CR. Hoever, hen used ith the target features, the derived shado features produce the best perforance of.3% CR. The reason for this is that the shado ask overlaps the target region and both and anage to include inforation that is already in u. The features on the other hand concentrate on the shado, giving relatively poor perforance on their on but giving the greatest iproveent hen cobined ith the target features, u. We ill use to represent the shado features in the folloing discussions. Table : Recognition error rates ith different shado features (%) Target u u+ u+ u+ BP BRD BTR BTR D T T ZIL3.8.4 ZSU S CR St.Dev We no present the recognition results based on 3 and 6 sequences, respectively. Table 2 shos the recognition error rate for each of the three feature sets for the to cases. In the 6 -sequence observation based recognition, the log likelihood of to adjacent test sequences are added and the recogniser chooses the one ith the highest su. We see that used on their on the u and paraeter sets give CR of.5% and.9% hich is in both cases an iproveent over the one-sequence results. When the to feature sets are cobined to for u + e find that the CR is reduced to.2% ith seven of the ten targets error-free. Table 2: Recognition error rates of to observations (%) 3 aziuth aperture 6 aziuth aperture Target u u+ u u+ BP BRD BTR BTR D T T ZIL3.8 ZSU S CR St.Dev or both the 3 -sequence and 6 -sequence tests, the addition of the features iproves the recognition rate for all targets except one, the BTR6 hich is considerably orsened. This indicates that, for this target at least, the assuption that the feature vector follos a ultivariate Gaussian distribution is a poor one. The error rate of.3% obtained using a 3 aperture can be directly copared ith other published results based on the STR database ith the sae recognition task. In [2], the authors obtained error rates of 4.% using an approach based on the SR iage and in [] an error rate of 7.8% as obtained hen perforing recognition on the HRR profiles directly. The STR database contains variants of the T72 tank and 3 variants of the BP2 vehicle, anifested by different realisations of the fuel tank, antenna, etc []. To evaluate the robustness of our recogniser to these variations, e trained odels on to BP2 variants, four T72 variants as ell as the eight other targets using the u + feature set. We then conducted recognition tests on all targets in the database using to iages. If the recogniser identified an incorrect variant of the correct tank odel, it as counted as an error in the Strict colun of Table 3 but as a correct identification in the Class colun. or targets included in the training set, the CR is.4 % and in ost cases the precise variant of a particular target is identified correctly. or the unseen variants listed in the loer section of the table, the CR as 8.6% ith ore than half of the errors arising fro the T72-s7 and T72-82 targets. The final colun of Table 3 shos the Class error rates reported in [] for the sae task. We see that for all targets except T72-s7 the recognition perforance for our proposed feature set is considerably better. We note hoever that since [] bases its recognition on individual HRR profiles, it ill be less sensitive to target otion than hen the feature set described here is used ith a large value of P. Table 3: Recognition error rates ith unknon targets (%) Target Strict Class Class fro [] BP2-c BP BTR7 4. BRD2 8.6 BTR D7 4.8 T62 9 T72-a T72-a T72-a T ZIL3.3 ZSU S 4.7 BP T72-a T72-a T72-a.5.9 T72-a T72-a T72-s T CONCLUSIONS This paper has presented a novel radar target recognition technique cobining to-diensional target and shado inforation. The ne technique copleents the SR-TR and HRR-TR techniques by using a feature extraction

6 ethod that is robust to noise and that can extract target and shado inforation accurately ith liited aziuth aperture length. Three ethods to obtain the shado features are investigated. The experiental results using STR database indicate that although they perfor relatively poorly on their on, the shado features derived using the ethod perfor exceptionally ell hen cobined ith USICderived target features. Using aziuth apertures of 3 and 6 in a -target classification task, they give overall classification error rates of.3% and.2% respectively. These results are considerably better than other published techniques using the sae dataset. 5. CKNOWLEDGEENT This research as supported by the UK od through ork funded by the Defence Technology Centre for Data and Inforation usion. REERENCES. L.. Novak, coparison of D and 2D algoriths for radar target classification, IEEE International Conference on Systes Engineering, 99, pp L.. Novak, S. D. Halversen, G. Oirka, and. Hiett, Effects of polarization and resolution on SR TR, IEEE Trans. on erospace and Electronic Systes, vol. 33, no., 997, pp J.. O Sullivan,. D. DeVore, V. Kedia, and. I. iller, SR TR perforance using a conditionally Gaussian odel, IEEE Trans. On erospace and Electronic Systes, vol. 37, no.,, pp D. P. Kottke,. Jong-Kae, and K. Bron, Hidden arkov odeling for autoatic target recognition, Proc. Thirty-irst siloar Conference on Signals, Systes and Coputers, vol.. 997, pp P. Runkle, L. H. Nguyen, J. H. cclellan, and L. Carin, ulti-aspect target detection for SR iagery using hidden arkov odels, IEEE Trans. On Geoscience and Reote Sensing, vol. 39, no.,, pp R. Bhalla, H. Ling, J. oore, D. J. ndersh, S. W. Lee, and J. Hughes, 3D scattering centre representation f coplex targets using the shooting and bouncing ray technique: a revie, IEEE ntennas and Propagation agazine, vol. 4, no. 5, 998, pp H. Chiang, R. L., oses, and L. C. Potter, odel-based classification of radar iages, IEEE Trans. on Inforation Theory, vol. 46, no. 5,, pp C. Nilubol, Q. H. Pha, R.. ersereau,. J. T. Sith, and.. Cleents, Hidden arkov odelling for SR autoatic target recognition, Proc. IEEE Int. Conf. coustics, Speech, Signal Processing, vol. 2, 998, pp R. Willias, J. Westerka D. Gross,. Paloion, and T. ister, utoatic target recognition of tie critical oving targets using D high range resolution (HRR) radar, IEEE Radar Conference, 999, pp D. E. Nelson, J.. Starzuk, and D. D. Ensley, Iterated avelet transforation and signal discriination for HRR radar target recognition, IEEE Trans. on Systes, an and Cybernetics, Part, vol. 33, no., 3, pp X. Liao, P. Runkle, and L. Carin, Identification of ground targets fro sequential high-range-resolution radar signatures, IEEE Trans. On erospace and Electronic Systes, vol. 38, no. 4, 2, pp B. Pei and Z. Bao, Radar target recognition based on peak location of HRR profile and Hs classifiers, Proc. Radar Conference, 2, pp T. D. Ross and J. C. ossing, The STR Evaluation ethodology, Proc. SPIE - The International Society for Optical Engineering, vol. 372, 999, pp T. Ross, S. Worrel, V. Velten, J. ossing, and. Bryant, Standard SR TR evaluation experients using the STR public release data set, Proc. SPIE - The International Society for Optical Engineering, vol. 337, 998, pp P. Lobardo,. Sciotti, and L.. Kaplan, SR prescreening using both target and shado inforation, Proc. IEEE Radar Conference,, pp J. Cui, J. Godnason and.brookes, utoatic Recognition of STR Targets using Radar Shado and superresolution eatures, is accepted by Proc. IEEE Int. Conf. coustics, Speech, Signal Processing, R. Schidt, ultiple eitter location and signal paraeter estiation, IEEE Trans. on ntennas and Propagation, vol., no. 3, 986, pp S. Kay, odern spectral estiation, st ed. Prentice Hall, P. Depster, N.. Laird, and D. B. Rubin, axiu Likelihood fro Incoplete Data via the E lgorith, Journal of the Royal Statistical Society, vol. 39, no., 977, pp S. Young, G. Everann, T. Hain, D. Kersha, G. oore, J. Odell, D. Ollason, D. Povey, V. Valtchev, and P. Woodland, The HTK Book, Cabridge University Engineering Dept, 2. [Online] vailable: 2. L.. Novak, utoatic target recognition using enhanced resolution SR data, IEEE Trans. on erospace and Electronic Systes, vol. 35, no., 999, pp BIOGRPHY Jingjing Cui as born in 98. She received her B. S. degree fro the Departent of Counication Engineering in udan University, China, in 2, and in 3 she received her Sc degree fro the Departent of Electrical and Electronic Engineering in Iperial College London, UK. She is currently doing her PhD degree at the Counication and Signal Processing Group in Iperial College. Her current research interests include echo cancellation and radar target recognition. Her contact address is: Counication and Signal Processing Grou Departent of Electrical and Electronic Engineering, Iperial College London, UK, SW7 2BT.