Signal processing techniques for landmine detection ...

2 downloads 0 Views 510KB Size Report
Nov 15, 2013 - GSI Helmholtz Centre for Heavy Ion Research View project · A.M. Zoubir · Technische Universität Darmstadt. 486 PUBLICATIONS 4,107 ...
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/3430977

Signal Processing Techniques for Landmine Detection Using Impulse Ground Penetrating Radar Article in IEEE Sensors Journal · March 2002 DOI: 10.1109/7361.987060 · Source: IEEE Xplore

CITATIONS

READS

86

1,208

5 authors, including: A.M. Zoubir

Christopher L Brown

Technische Universität Darmstadt

Opel Automobile

486 PUBLICATIONS 4,107 CITATIONS

28 PUBLICATIONS 265 CITATIONS

SEE PROFILE

SEE PROFILE

Braham Barkat Petroleum Institute (UAE) 73 PUBLICATIONS 824 CITATIONS SEE PROFILE

Some of the authors of this publication are also working on these related projects: Development of PD Diagnostic System for Power Cable Terminations based on PD-emitted Electromagnetic Waves View project GSI Helmholtz Centre for Heavy Ion Research View project

All content following this page was uploaded by Braham Barkat on 15 November 2013. The user has requested enhancement of the downloaded file.

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

41

Signal Processing Techniques for Landmine Detection Using Impulse Ground Penetrating Radar Abdelhak M. Zoubir, Senior Member, IEEE, Ian James Chant, Christopher L. Brown, Member, IEEE, Braham Barkat, and Canicious Abeynayake

Abstract—Landmines are affecting the lives and livelihoods of millions of people around the world. A number of detection techniques, developed for use with impulse ground penetrating radar, are described, with emphasis on a Kalman filter based approach. Comparison of results from real data show that the Kalman filter algorithm provides the best detection performance, although its computational burden is also the highest. Index Terms—Buried object detection, Kalman filtering, radar signal processing.

I. INTRODUCTION—THE PROBLEM

L

ANDMINES and unexploded ordnance (UXO) are a legacy of war, insurrection, and guerilla activity. Landmines kill and maim approximately 26,000 people annually. Direct casualties are not the only problem. In Cambodia, whole areas of arable land cannot be farmed due to the threat of landmines. United Nations relief operations are made more difficult and dangerous due to the mining of roads [1]. Current demining techniques are heavily reliant on metal detectors and prodders. In many circumstances, the prodder is the first, and in all cases, the last resort. The advent of nondisturbance fused mines makes prodding a dangerous operation. Mechanical devices such as ploughs, rollers, and flails are usually followed by manual demining to obtain the desired level of clearance. These machines are expensive for developing countries. Dogs are good when they work but can only operate for limited periods and must be acclimatized. In order to assist deminers, a range of advanced sensor technologies are being investigated, including • metal detectors (MD)—capable of finding even low-metal content mines in mineralized soils [2]; • nuclear magnetic resonance, nuclear quadrapole resonance (NQR), fast neutron activation and thermal neutron activation—detect the presence of the explosive material in landmines [3], [4]; Manuscript received September 8, 2000; revised November 13, 2001. The associate editor coordinating the review of this letter and approving it for publication was Dr. Thaddeus A. Roppel. A. M. Zoubir is with the Curtin University of Technology, Perth, Australia (e-mail: [email protected]). I. J. Chant and C. Abeynayake are with the Defence Science and Technology Organization, Salisbury, Australia. C. L. Brown was with the Curtin University of Technology, Perth, Australia. He is now with the Department of Signals and Systems, Chalmers University of Technology, Göteborg, Sweden (e-mail: [email protected]). B. Barkat was with the Curtin University of Technology, Perth, Australia. He is now with Nanyang Technological University, Singapore. Publisher Item Identifier S 1530-437X(02)02153-X.

Fig. 1. ImGPR system.

• thermal imaging (TI) and electrooptical (EO) sensors—detect evidence of a buried object, such as disturbed ground or the thermal effect of having a mine just below the surface [5], [6]; • biological sensors such as dogs, pigs, bees, and birds [7]; • chemical sensors, such as thermal fluorescence and chromatographic techniques—detect airborne and water borne presence of explosive vapors [8]. In this discussion, we will concentrate on Ground Penetrating Radar (GPR) [9]. This ultrawide band radar provides centimeter resolution to locate even small targets. GPR operates by detecting the dielectric contrasts in the soils, which allows it to locate even nonmetallic mines. Unfortunately, this technology can suffer false alarm rates as high as that of metal detectors. There are two distinct types of GPR, time-domain and frequency domain. Time domain or impulse GPR transmits discrete pulses of nanosecond duration and digitizes the returns at GHz sample rates. Frequency domain GPR systems transmit single frequencies either uniquely, as a series of frequency steps, or as a chirp. The amplitude and phase of the return signal is measured. The resulting data is converted to the time domain. As with metal detection, GPR automatic detection [10] and classification algorithms are being developed. In this discussion, we deal with buried anti-tank (AT) and anti-personnel (AP) landmines which require close approach or contact to activate. Trip wire activated area mines and newer off-route or side-attack mines are a danger that has not been specifically addressed here. II. GPR FOR LANDMINE DETECTION: OVERVIEW OF SYSTEM AND DATA COLLECTION A series of measurements has been taken using a set of targets buried in various types of soil. An FR-127-MSCB impulse

1530–437X/02$17.00 © 2002 IEEE

42

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

Fig. 2.

ImGPR unit running over a sandbox.

GPR (ImGPR) system developed by the Commonwealth Scientific and Industrial Research Organization (CSIRO), Australia, has been used for these measurements [11]. The system collects 127 returns, or soundings, per second, each composed of 512 samples with 12 bit accuracy. The sounding range may vary from 4 ns to 32 ns. The GPR system uses bistatic bow-tie antennas which transmit wideband, ultrashort duration pulses (see Fig. 1). A dielectric anomaly in the soil may cause the signal to be reflected back to a separate receiver antenna. This information is converted from nanoseconds to milliseconds so that it may be digitized by a conventional A/D converter for processing and display. The center frequency and bandwidth of the transmitted pulse can be varied by changing the antenna and are chosen with respect to the required depth of penetration, soil type and size of the object to be detected. In this experiment, we used antennas with a center frequency 1.4 GHz and 80% bandwidth. The GPR unit is suspended above the ground surface at a height of between 0.5 to 2 cm. Its motion is controlled by a stepper motor unit running along a track at a constant velocity, as shown in Fig. 2. Since the motion of the GPR is controlled by a stepper motor, with constant speed, running on a straight track, these samples correspond to distances from the starting point of the run. The measurements form a two dimensional matrix, referred to as a radargram or B scan, and is used for visual inspection of the data on the acquisition computer and in laboratory analysis. A sample radargram showing two targets at approximately 55 cm and 100 cm is displayed in Fig. 3. A return at a certain position along the distance axis is called an A scan. An example of A scans in the presence and absence of a surrogate mine are displayed in Fig. 4. Some of the targets used in the trial are listed in Table I. Three of the targets were surrogate land-mines developed by the Defence Science and Technology Organization (DSTO) countermine research project, Salisbury, South Australia [11], [12]. These targets, referred to as ST-AP(1), ST-AP(2), and ST-AP(3), are surrogate AP mines modeled after the M14,

Fig. 3. Radargram showing target positions.

Fig. 4. A scans in the presence (dashed) and absence (solid) of a surrogate mine in clay soil.

PMN, and PMN-2 blast mines, respectively. The PMN and PMN-2 are AP mines with nonmetallic casings. The M14 is an AP mine with almost no metal content and small size. As such, it is a very difficult target to detect. III. SIGNAL PROCESSING METHODS A. Kalman Filtering Method 1) Principle: The radargram data is divided into nonoverlapping horizontal strips which correspond to layers of approximately constant depth. In general, one trace of the measured GPR data can be modeled as [13] (see the equation at the bottom denotes the signal time of the page). where

target-free model target-present model

(1)

ZOUBIR et al.: SIGNAL PROCESSING TECHNIQUES FOR LANDMINE DETECTION USING IMPULSE GROUND PENETRATING RADAR

43

TABLE I MINE-LIKE SURROGATE TARGETS USED

TABLE II KALMAN FILTER EQUATIONS FOR TARGET-FREE AND TARGET-PRESENT CASES. THE ROWS CONTAIN (FROM TOP TO BOTTOM) STATE SPACE REPRESENTATION, MEASUREMENT EQUATION, STATE VECTOR AND UPDATE EQUATIONS, RESPECTIVELY

samples at distance of the receive antenna, that is, the distance of the considered trace from the origin. The target signal, background signal, and interference are repand , respectively. If the resented by radargram is divided into horizontal strips where each strip has a height equal to , we define the measurement vector as with

and is the largest integer . a) Target-free model: Following the random walk model for the background only case, we define the qui, and corresponding measurement escent state, in Table II, for and equation, . Here, the -vector valued and correspond, respectively, to the process noise and measurement noise. They are both independently and identically distributed (iid) Gaussian vectors.

A set of Kalman filters is used to estimate the background and target signals from noisy observations. In the quiescent , written in predictor-corphase, the vector is estimated as rector format, using a set of equations given in Table II. In these is an identity matrix, is an equations, a priori error covariance matrix, is the Kalman is the updated error covariance matrix, filter gain, is the process noise covariance matrix. The error and and covariance matrix is initialized to zero, i.e., , where the process covariance matrix to is the initial estimate of the variance of the process . noise sequence b) Target-Present Model: If a target is present, the and , as state space representation contains for well as the corresponding measurement equation . The vector is defined in the same

44

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

way as and the vector is a random bias accounting for the changes in the target signal. Note that, in this case, the background is kept constant and equal to the value estimated prior to target detection. When a target is present, we estimate the augmented state as given in vector where Table II along with the update equations for

The augmented state is initialized at the trace (or distance) by setting the bias to zero and the error covariance matrix to

2) Detection: In the quiescent state, the algorithm uses a test based on the measurement prediction error (or innovation) to detect the position of a possible target. The target is characterized by a “large” innovation. The detection model for a fixed trace is as follows for H

• This proposed target algorithm cannot determine the end of the target. However, for the purposes of evaluation here, we assume that the width of the target is known. This information is used to re-activate the lower state model. • For noisier backgrounds, it is suggested to compute the detection statistic as a moving sum of NIS over a sliding window of length as . 3) Background Adaptation: In order to account for natural variations in the background, the noise covariance matrix must be continually updated, when in the quiescent . The state model only, through is determined by comparing the NIS scaling factor to a set of thresholds, as described in Table III where . This approach was motivated by the continuous noise level adjustment technique presented in [14]. 4) Dependency Between Layers: Although in the formulation of the above procedure, it was assumed that a target may be present in only one of the layers, the true effect of a landmine return may not be confined to one section of the GPR return. That is, the presence of a target is likely to trigger detection decisions in more than one of the hypothesis tests. Similarly, false alarms should occur independently in the tests, and, in the case of the usual “statistical” false alarms present in hypothesis testing, they do. However, physical soil anomalies may trigger effects visible in multiple layers, as for landmine targets. As a result, the multiple hypotheses are not independent, and though the nominal levels of each of the individual tests may be specified, the nominal global level of significance cannot.

K For a given trace , we compute the measurement prediction , the updated measurement prediction covariance error , and the normalized innovation squared (NIS) matrix

B. Background Subtraction Perhaps the most common and basic form of target detection uses the model depicted in Fig. 5. This model translates to the analytic form (2)

Under the null hypothesis H are distributed with degrees of freedom (recall that is the sum of squared independent Gaussian variables). For a fixed trace , each strip is tested using the test with the significance level so that

In the case when H is rejected for at least of the total strips, the null hypothesis is rejected for the whole trace. A target consecutive GPR traces the is declared present if for at least null hypothesis is rejected. In that case, we decide to change to, or activate, the augmented state model. When the target detection is declared at a position , the augmented state is initialized and is given by , where and at are suitably chosen constants. c) Notes: represents the spatial position where detec• If , we assume that tion first occurred. If we choose the target started before it was first detected.

The objective is to test the presence or absence of a target, that is, against K . Under H, the difference between H the estimated signal and the target-free signal (or background) will be small, while under K, there will be a difference due to the return from the buried target. The target detection procedure tests the significance of this difference, using amplitude detecwhere tion. This suggests the statistic, is an estimator of the background signal. The hypothesis H is larger than a threshold, , where is is rejected if obtained empirically. relies on the assumption that background The estimator clutter has slow spatial variation, and, therefore, can be adaptively estimated from previous traces in several ways. • Averaging across all traces—a quick estimate since it is only calculated once for a data set. Since it can be assumed that even under H there will be some variation in the background return between the start and the end of the recordings, it may not be very accurate. Only in tightly controlled, homogeneous environments may these differences be assumed to be negligible.

ZOUBIR et al.: SIGNAL PROCESSING TECHNIQUES FOR LANDMINE DETECTION USING IMPULSE GROUND PENETRATING RADAR

45

TABLE III BACKGROUND ADAPTATION FOR THE KALMAN FILTER METHOD

Fig. 5. Received signal x(t) as a superposition of background return, b(t), and scaled target signal,  1 s(t).

• The median across all traces—is also susceptible to spatial variations, though the median may provide a more accurate measure where the data is skewed. • Running means or medians—by choosing an appropriately wide moving window, an estimate of the background is formed by taking the mean or median of the surrounding traces. The window must be wide enough to allow an accurate estimate to be made, with low variance, while narrow enough to avoid introducing effects from the changes in the local background characteristics. C. Matched Filter Deconvolution By considering the wave propagation path through the soil as a channel or filter, another signal model can be used. Consider a convolutive relationship, as shown in Fig. 6. This can be written as H

(3)

where

and is the channel or filter impulse response. The input is either H or K . The signal , which indicates the presence of the target, is clearly different to due to the different model. Under H and, therefore, is simply the observed response from background only. The is equivalent to the background estimation estimation of process used for background subtraction. by an appropriate By deconvolving the observations, , amplitude detecestimate of the background signal, using the test statistic tion can be performed on

Fig. 6.

Channel model for matched filter deconvolution.

deconv indicates the deconvolution operation.

where deconv

D. Wavelet Packet Decomposition Due to the transient nature of the backscattered waveform, it is natural to investigate wavelet based techniques in the analysis of ImGPR returns. However, due to the variation from trace to trace in the relative positions of the receiver antenna and the buried targets, time shifts in the corresponding returns can be expected. As a result, we suggest the so-called Translation Invariant Wavelet Packet Decomposition (TIWPD) [15]. Briefly, TIWPD is achieved by searching for the best basis that minimizes an entropy based cost function. At each level of resolution, the signal is decomposed and an entropy-based cost function evaluated. This is compared to the equivalent cost function when the signal is shifted, to find the best basis. The wavelet based detection procedure [16] consists of the steps in Table IV. The distribution of the test statistic used here remains unresolved, although it appears Gaussian under H. We recommend that the median of the test statistics under H be estimated from traces known or assumed to be target-free [16]. This median is multiplied by an appropriate constant, , to set the threshold. is an appropriate Experimentally, it has been found that value. E. Trimmed Average Power The idea behind the trimmed mean approach is that the existence of a target is indicated by a change in the “average” signal power from one trace to another [17]. The average power of a signal (i.e., ImGPR trace) is estimated using the trimmed mean of the periodogram.

46

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

TABLE IV DETECTION USING TRANSLATION INVARIANT WAVELET PACKET DECOMPOSITION

It was found that the amount of right trimming was far more critical to performance than left trimming. This was expected since a target should increase the power of the return and the periodogram ordinates are positive and skewed to the right. Therefore the ordinates removed by the effect of left trimming will be of low magnitude and relatively unaffected by the presence of a target. From testing over a number of target scenarios, soil types and nominal levels of significance, right trimming of 5% was found to be best and for the remainder of the paper, trimming of 5% is used. Optimal trimming parameters may be determined through a bootstrap procedure to minimize estimator variance [17]. IV. COMPARISON AND DISCUSSION Our comparison of the signal processing techniques described above will be based on detection rates, false alarm rates, and computational complexity. To do this, we use receiver operating characteristic (ROC) curves and computational time [18]. A. ROC Evaluation Due to the large number of parameters in the detection algorithms, the comparison of the various detection techniques is not a straight-forward task. A commonly used method to com-

pare detector performance is through ROC curves. The rates of false alarm and correct detection are found for varying nominal levels of significance. ROCs are considered to be a good way to compare detectors as they incorporate both these performance indicators. In using the available data sets, the unknown ground truth and spatial variation of the background create difficulties. Strictly speaking, it is not possible to simply run the detectors over target-free or target-present data and estimate the probabilities of false alarm and correct detection as the average number of detections. There is a correlation between the background component of traces, and, therefore, a correlation between the detection decisions, unless it can be asserted that the effect of the background is completely removed. Several techniques can be used for background estimation, as discussed in Section III B. Here, a moving window of background traces is used for background estimation. A test area is manually identified from the target-present . The same area of recordings to find the detection rate, ground is then tested in the target-free recordings to find the . The testing procedure is described in false alarm rate, Tables V and VI. The threshold setting area in step 1 of Table V was chosen to start at 150 traces before the start of the test area and to finish 50 traces before the start of the test area for the results to be shown here.

ZOUBIR et al.: SIGNAL PROCESSING TECHNIQUES FOR LANDMINE DETECTION USING IMPULSE GROUND PENETRATING RADAR

47

TABLE V ESTIMATING PROBABILITIES OF FALSE ALARM AND CORRECT DETECTION FOR ROCS

TABLE VI ADAPTIVE/MOVING WINDOW BACKGROUND ESTIMATE

The assumption in step 1 of Table V may not be strictly correct, since, although the test area definitely contains a target, the area immediately before it may still contain some effects from the target too. Here, we assume that these effects are too small to trigger a detection. If this is not the case, this implies that the detection of the target has occurred on the edge of the target rather than in the test area. Thresholds are set from a small target-free area. If recordings of larger, target-free areas were made, this may be expanded. For the testing here, 100 traces were used for threshold setting. Therefore, some of the resulting ROCs may not seem smooth. ROCs are presented in Figs. 7 and 8. All targets were buried at a depth of 5 cm. We choose background estimation lengths and traces. For each technique, the best of

estimation length was found for each soil type and used in the ROCs shown. It can be stated that the Kalman filter based technique appears to have the best overall performance by a significant margin. Although the ROCs do not show that it is the best for all scenarios, it does not fail under any of them. All other detectors can be shown to perform poorly in one or more cases—in particular ST-AP(3) in loam, where the probability of detection is often lower than the probability of false alarm. This is worse than an equal chance decision and is highly undesirable as it indicates a complete failure of the detector. Deconvolution appears to also have good performance, except for ST-AP(1) in clay, but is probably the best for ST-AP(2) and ST-AP(3). The remaining three detection methods achieve

48

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

Fig. 7. ROC of ST-AP(1) (left), ST-AP(2) (center), and ST-AP(3) (right) targets in clay soil. Note the breakdown of some detectors.

Fig. 8. (Left) ROC of ST-AP(1), (center) ST-AP(2), and (right) ST-AP(3) targets in loam soil. Note the breakdown of some detectors for ST-AP(3). Also note that all techniques are showing detection rates of 1 or close to it for ST-AP(2), hence the change in ordinal scale. The same legend is valid for all three figures; however, it has been removed from the right-most figure as it obscured the plot. TABLE VII NUMBERS OF TRACES USED FOR BACKGROUND ESTIMATION TO YIELD BEST RESULTS IN CLAY

similar overall performance. Due to variability in results across the scenarios, we continue the analysis by considering subsets of the presented results. 1) Clay Soil: While the Kalman filter based approach was the most successful in clay, the trimmed average power was seen to be the next best. Deconvolution proved to be ineffective in detecting the ST-AP(1) targets, yet not for ST-AP(2) or ST-AP(3). 2) Loam Soil: Again, the Kalman filter yielded the best detection. The ST-AP(1) and ST-AP(2) mines were more detectable than in clay, though the opposite can be said for ST-AP(3). The wavelet and background subtraction were again very similar; however, the former may be said to be slightly better due to its false alarm performance, as will be discussed later. 3) Background Estimation Length: Some detectors appear to be more sensitive to the number of traces used to estimate the

background. Tables VII and VIII give recommended lengths that produce the best detection results for different scenarios. The entries under the heading “Overall” are those that were used in the ROCs shown above. Testing indicates that little difference can be seen for the Kalman Filter detector with varying , since detection rates were often very close to 1. No single background estimation length can be recommended for all cases, although the Deconvolution method appeared to vary the least—taking on “optimal” values of either 10 or 30. By contrast, the best values for the trimmed average power method varied wildly. If assessments of the relative performance of the techniques are made on the best results for each detector for each scenario, the Kalman filter is still the best. The trimmed average power is the next, rather than deconvolution.

ZOUBIR et al.: SIGNAL PROCESSING TECHNIQUES FOR LANDMINE DETECTION USING IMPULSE GROUND PENETRATING RADAR

49

TABLE VIII NUMBERS OF TRACES USED FOR BACKGROUND ESTIMATION TO YIELD BEST RESULTS IN LOAM

Fig. 9. Probability of false alarm (P ) versus nominal level of significance ( ) of a (left) ST-AP(1), (center) ST-AP(2), and (right) ST-AP(3) targets at a depth of 5 cm in clay. The same legend applies to all figures.

Fig. 10. Probability of false alarm (P ) versus nominal level of significance ( ) of (left) ST-AP(1), (center) ST-AP(2), and (right) ST-AP(3) targets at a depth of 5 cm in loam. The same legend applies to all figures.

4) Nominal Level of Significance and Probability of False Alarm: As noted previously, the nominal level of significance cannot be set for the Kalman filter based technique due to the correlation between the multiple hypotheses. Plots for the remaining techniques, showing the relationship between the nom, are shown inal level, , and the observed false alarm rate, below in Figs. 9 and 10. The false alarm rate should not be dependent on the target under consideration. However, we notice significant variations here. Different test areas, corresponding to different physical areas of ground, were considered for different target cases. This is the cause of the variation observed. A general statement can be made that the level is better maintained in clay, especially at the lower levels of . This would be the operating region of a landmine detector. Exceeding the

nominal level appears to be more prevalent at higher levels. The wavelet based detector appears to be best at maintaining the level, while the trimmed mean power and deconvolution methods often fail at quite low levels. In loam, failure to maintain the level may indicate anomalies in the soil at these locations, especially when using the test region corresponding to the ST-AP(2) and ST-AP(3) cases. This view is strengthened by noting that the test region used to find the false alarm rates for the ST-AP(1) case was far from the others. B. Computational Load Shown in Table IX are the number of flops required by the and using a backMatlab code used for the evaluation of . ground estimation length of

50

IEEE SENSORS JOURNAL, VOL. 2, NO. 1, FEBRUARY 2002

TABLE IX COMPUTATIONAL COST FOR THE EVALUATION OF ONE (P ; P ) PAIR USING A BACKGROUND ESTIMATION LENGTH OF FIVE TRACES AND TEST AREA WIDTH OF 50 TRACES

The Kalman filter based approach can be seen to be very computationally expensive compared to the other methods. The next most demanding technique, deconvolution, is approximately 20 times faster. Not surprisingly, background subtraction has the lightest load.

form the problem toward system identification with known input signal model. The results shown in this paper cover a significant subset of the available targets and soils. However, significant variations in the results were observed for the target and soil combinations tested. This made it difficult to draw clear conclusions in some cases. Different methods of evaluating ROCs may yield different results and the extrapolation of these conclusions to other targets and soils may not be valid. It is felt that the detection performance comparison framework developed here will greatly assist in the refinement, development and extension of existing and new detectors for the detection of landmines. Similar detection algorithms need to be developed and tested for other sensors. A fused system using multiple techniques and sensors may, ultimately, result. ACKNOWLEDGMENT The authors would like to acknowledge the significant contributions of Dr. D. Carevic, especially in the development of the Kalman filtering detection algorithm. REFERENCES

V. CONCLUSION We have presented some signal processing techniques for use with GPR and compared their detection performance using receiver operating characteristic curves. We have found that the way the background is estimated and the number of traces considered in this estimation has an impact on the analysis results. An adaptive background estimate yields the best results in terms of reducing the probability of false alarm. An analyst should consider plotting the ROC curves for all techniques for the particular environment they are working in (type of soil, type of target, depth, target size, etc.) before deciding on the best algorithm to use. It is seen from the ROCs shown here that the Kalman filter based detector provides the best overall performance. It may be significant to note that the Kalman filter approach explicitly incorporates the background component of the signal return into its model, while the others are essentially different methods of detecting a change in a background-adjusted trace. The cost of the superior performance of the modified Kalman filter approach is the substantial increase in computational load. While significant time savings may be made through optimization and the parallelization of the Kalman filter code, it will remain a more complicated approach than the background subtraction, deconvolution, trimmed average power and, to a lesser extent, wavelet based detectors. A multiple test procedure may be incorporated into the Kalman filter approach in the future to allow the setting of the nominal global level of significance. After the Kalman filter, the trimmed average power appears to achieve good detection performance for a relatively light computational load. The encouraging results that flow from trimming may lead to its application in the calculation of different test statistics to improve performance. The deconvolution technique could be improved by incorporating modeling into the system and parameterization of the input signal. This will trans-

[1] The Canadian Forces Mine Database 99, 1999. [2] L. Collins, P. Gao, S. Tantum, J. Moulton, L. Makowsky, D. Reidy, and R. Weaver, “A comparison of statistical signal processing algorithms for detection and identification of low metal mines,” presented at the UXO/Countermine Forum, Anaheim, CA, May 2000. [3] A. D. Hibb, G. A. Barrall, P. V. Czipott, D. K. Lathrop, Y. K. Lee, E. E. Magnuson, R. Matthews, and S. A. Vierkotter, “Landmine detection by nuclear quadrapole resonance,” in SPIE Conf. Detection Remediation Technologies Mines, Minelike Targets III, Orlando, FL, Apr. 1998, pp. 522–532. [4] R. A. Craig, A. J. Peurrung, and D. C. Stromswold, “Mine detection using timed neutron moderation,” presented at the UXO/Countermine Forum Conf. Proc., Anaheim, CA, May 2000. [5] T. Miller, “A mathematical model of heat conduction in soil applied to thermal detection of landmines,” presented at the Proc. Australian-Amer. Joint Conf. Mine Warfare, MINWARA’99, Sydney, NSW, Australia, July 1999. [6] M. Lundberg, “Infrared land mine detection by parametric modeling,” Proc. IEEE Int. Conf. Acoustics, Speech, Signal Processing, ICASSP 2001, May 2001. [7] Sci/Tech-Making a Beeline for Mines [Online]. Available: http://news.bbc.co.uk/hi/english/sci/tech/newsid_373 000/373 684.stm [8] M. Fisher, C. Cumming, M. Fox, M. la Grone, S. Jacob, D. Reust, M. Rockley, and E. Towers, “Sensing ultra-trace concentrations of landmine chemical signature compounds in the air over landmines using a manportable chemical sniffer,” presented at the UXO/Countermine Forum Conf. Proc., Anaheim, CA, May 2000. [9] D. J. Daniels, Surface-Penetrating Radar. London, U.K.: Inst. Elect. Eng., 1996. [10] J. Groenenboom and A. G. Yarovoy, “Data processing for a landmine detection dedicated GPR,” presented at the Proc. Eighth Int. Conf. Ground Penetrating Radar, GPR2000, Gold Coast, Australia, May 2000. [11] I. J. Chant and A. R. Rye, “Overview of current radar land mine detection research at the Defence Science and Technology Organization, Salisbury, South Australia,” Proc. IEE Conf. Detection Abandoned Land Mines, Oct. 1996. [12] B. C. Y. Wong, I. J. Chant, G. N. Crisp, K. Kappra, K. Strugess, A. R. Rye, and K. Sherbondy, “Suggested soil characterization techniques and surrogate targets for ultra-wide-band radar mine detection experiments,” presented at the Proc. SPIE AeroSense’97, Orlando, FL, 1997. [13] D. Carevic, “Kalman filter-based approach to target detection and targetbackground separation in ground-penetrating radar data,” in SPIE Conf. Detection Remediation Technol. Mines, Minelike Targets IV, 1999, pp. 1284–1288. [14] Y. Bar-Shalom and X.-R. Li, Estimation and Tracking: Principles, Techniques and Software. Norwood, MA: Artech House, 1993.

ZOUBIR et al.: SIGNAL PROCESSING TECHNIQUES FOR LANDMINE DETECTION USING IMPULSE GROUND PENETRATING RADAR

[15] D. Carevic, “Clutter reduction and target detection in ground penetrating radar data using wavelets,” in SPIE Conf. Detection Remediation Technologies Mines, Minelike Targets IV, 1999, pp. 973–978. , “Clutter Reduction and Detection of Minelike Targets in Ground [16] Penetrating Radar Data Using Wavelets,” Defence Science and Technology Organization, Salisbury, Australia, Tech. Rep. DSTO-TR-0893, Dec. 1999. [17] A. M. Zoubir, D. R. Iskander, I. Chant, and D. Carevic, “Detection of landmines using ground-penetrating radar,” in SPIE Conf. Detection Remediation Technologies Mines, Minelike Targets, IV, 1999, pp. 1301–1312. [18] A. M. Zoubir, B. Barkat, C. L. Brown, and I. Chant, “A comparison of some landmine detection and estimation techniques using ground penetrating radar,” in Proc. 10th Australasian Remote Sensing Photogrammetry Conf., vol. 1, Adelaide, Australia, Aug. 2000, p. 785.

Abdelhak M. Zoubir (SM’97) received the Dipl.-Ing. degree (B.Sc./B.Eng.) from Fachhochschule Niederrhein, Niederrhein, Germany, in 1983, the Dipl.-Ing. (M.Sc./M.Eng.) and the Dr.-Ing. (Ph.D.) degree from Ruhr University Bochum, Bochum, Germany, in 1987 and 1992, all in electrical engineering. Early placement in industry (Klockner-Moeller & Siempelkamp AG) was then followed by Associate Lecturership in the Division for Signal Theory at Ruhr University Bochum. In June 1992, he joined Queensland University of Technology, Brisbane, Australia, where he was Lecturer, Senior Lecturer, and Associate Professor in the School of Electrical and Electronic Systems Engineering. In March 1999, he became Professor of Telecommunications in the Australian Telecommunications Research Institute (ATRI) and School of Electrical and Computer Engineering at Curtin University of Technology, Perth, Australia, where he is Head of the Department of Electronics & Communications Engineering. His general interest lies in statistical methods for signal processing with applications in communications, radar, sonar, biomedical engineering, and vibration analysis. His current research interest lies in bootstrap techniques for modeling nonstationary and non-Gaussian signals. Dr. Zoubir was the Technical Chairman of the 11th IEEE Workshop on Statistical Signal processing held in Singapore in August 2001. He is a member of the IEEE Technical Committee on Signal Processing Theory and Methods and an Associate Editor of the IEEE Transactions on Signal Processing. He is also a member of the Institute of Mathematical Statistics.

Ian James Chant was born in Brisbane, Australia. He received the B.Sc. and B.Sc. (Hon.) degrees, with concentration areas in physics, from Griffith University, Brisbane, in 1979 and 1980, respectively. In 1988, he received the M.Sc. from Queensland University, Brisbane, and, subsequently, the Ph.D. degree in 1992 for his work in developing a magnetotelluric deep sounding system and new analysis methods for magnetotelluric signals. He subsequently worked as a research assistant at Macquarie University, Sydney, Australia, helping to develop a magnetotelluric geophysical prospecting system under a NERDEC grant. Since 1992, Dr. Chant has worked as a Research Scientist for the Defence Science and Technology Organization (DSTO), Salisbury, Australia. The majority of his work for DSTO has been in the area of landmine detection and remediation. He is a committee member for the land section of the Mine Warfare conference (MINWARA) and for the SPIE AeroSense, Detection and Remediation Technologies for Mines and Mine-Like Targets.

View publication stats

51

Christopher L. Brown (M’01) received the B.Eng./B.InfoTech. degree from Queensland University of Technology, Brisbane, Australia, in 1995. He subsequently received the Ph.D. degree in electrical engineering from Curtin University of Technology, Perth, Australia, in 2000. He has held post-doctoral research fellowships at Curtin University, Perth, Australia, and, since May 2001, Chalmers University of Technology, Göteborg, Sweden. His research interests lie in statistical signal processing and, in particular, detection and goodnees-of-fit problems in impulsive noise, as well as the application of signal processing methods to landmine detection.

Braham Barkat received the degree of “Ingenieur d’Etat” in electronics in 1985 from the National Polytechnic Institute of Algiers (ENPA), Algiers, Algeria. In 1986, he joined the University of Colorado, Boulder, where he received the M.S. degree in control systems in 1988. In 1989, he joined the University of Blida, Blida, Algeria, where he held a lecturer position in digital and advanced control systems. In 1996, he joined the Signal Processing Research Centre at Queensland University of Technology (QUT), Brisbane, Australia, as a Senior Research Assistant and then as a Ph.D. candidate in signal processing. From September 1999 to November 2000, he was a Postdoctoral Research Fellow, first at QUT and then at Curtin University of Technology, Perth, Australia. In November 2000, he joined the School of Electrical & Electronic Engineering at Nanyang Technological University, Singapore, as an Assistant Professor. His research interests include time-frequency signal analysis, estimation and detection, statistical array processing, and signal processing in telecommunications.

Canicious Abeynayake was born in Sri Lanka. He received the M.S. and Ph.D. degrees in physics from the Friendship University, Moscow, Russia, in 1984 and 1989, respectively. From 1989 to 1995, he was employed as a senior lecturer at the Rzeszów Pedagogical University, Rzeszów, Poland. From 1995 to 1999, he was engaged in signal processing-related research projects at the Co-operative Research Centre for Sensor Signal and Information Processing, Adelaide, Australia, and the Flinders University of South Australia, Adelaide. Since 1999, he has been a Research Scientist at the Defence Science and Technology Organization, Salisbury, Australia. His research interests include signal and image processing, data fusion, and estimation theory.