A Discrimination Method for Landmines and Metal Fragments Using Metal Detectors

by Alex M. Kaneko, Edwardo F. Fukushima and Gen Endo [ Tokyo Institute of Technology ]
- view pdf

While discrimination methods for distinguishing between real mines and metal fragments would greatly increase the efficiency of demining operations, no practical solution has been implemented yet. A potentially efficient method for the discrimination of metallic targets using metal detectors uses a high-precision robotic manipulator to scan the minefield. Further field research is needed, however, before this method can deploy for operational use.

Current detection and clearance methods suffer from high false-alarm rates (FAR) and are costly, dangerous and time consuming. In 2001, the Tokyo Institute of Technology began work on a semi-autonomous mobile robot, the Gryphon (Figure 1), to facilitate the mine-detection process.1 The robot’s manipulator is equipped with tools for cutting vegetation and uses mine sensors to scan rough terrain, record data and note suspect locations by marking the ground. During experiments in test fields of flat terrain with no vegetation, the Gryphon proved as efficient as human operators when using a mine detector based on electromagnetic induction, such as a metal mine detector (MMD).2 The Gryphon proved superior when compared to human operators in terms of reducing FAR and increasing probability of detection. However, similar to other demining solutions, FAR is still problematic with the Gryphon.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.

Figure 1. The demining robot Gryphon and its metal mine-detector signal visualization.
All figures courtesy of the authors.

Problem Statement

One of the greatest problems in manual humanitarian landmine detection and removal involves high FAR, which are inherent to the use of electromagnetic induction-based detectors. Currently, no commercially available MMDs can distinguish landmines from other metal fragments. Some electromagnetic induction-based detectors, however, can select metal types to be searched, such as gold detectors.3 Similarly, MMDs can be used for the discrimination of landmines and other metal fragments, as shown by research in the following topics:

  1. Algorithms for evaluation of detected signals using models of physical phenomena4,5,6
  2. Feature extraction from MMD signals and classification of data according to metal type, size or depth of the metal fragments7,8,9
  3. Algorithms that combine time domain analysis and frequency domain analysis10,11

Some methods also rely on a dual-sensor approach, which combines two sensors and an MMD with ground-penetrating radar (GPR).12,13 However, a high level of expertise is still needed to properly evaluate the obtained data (image or sound). Moreover, discrimination has a large safety margin, which keeps FAR high. Another interesting method that has been reported uses image processing, MMD-signal surface area and volume calculation to estimate size and material, followed by depth estimation, which is achieved by placing the MMD at different angles.9 Despite reducing FAR to 39%, this method requires too much additional information from several depths (layers) besides the standard scan for discrimination, which considerably slows the demining operation by many minutes.8

Unfortunately, these methods have yet to be successfully implemented for use in practical demining tasks. Here, preliminary research on a potentially faster, newer, more accurate, on-site method (no need for additional scans) for discrimination of metallic targets using metal detectors is presented, and takes advantage of high-precision scans of the minefield using a robotic manipulator as shown in Figure 1.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.

Figure 2. SRMMDS for different targets and different postures and depths.

 

Robotic Scanning and Sensor Data

The usual scanning procedure consists of manually swinging the MMD sideways while advancing the search head in increments between one scan and another. A robotic arm, which achieves higher precision and repeatability, can conduct a similar procedure. For a human deminer, the MMD signals (called V[%] here) are transformed into sound, and the deminer must remember and search the position of the ground target. For a robotic system, the sensor signal can be transmitted to a computer and easily associated with the location of the manipulator. The signal can be processed in real time, and the user can easily visualize it (Figure 1). For the Gryphon system, the target position can be marked directly on the ground by painting or placing colored markers on the spot.14

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.

SRMMDS uniqueness. Figure 2 shows a 3-D plot of the MMD signal, also known as a spatially represented metal mine-detector signal (SRMMDS). SRMMDS drastically changes according to postures and target types. Depending on the target, SRMMDS will present different characteristics, which can include physical properties such as depth, material, posture, shape, size and soil conditions. This implies that if a database of SRMMDS for every target in every condition could be prebuilt, one would only need to compare the SRMMDS obtained in the minefield to get the closest match in the database, which would identify the target, as well as the target’s depth and posture. Even though some metal detectors can discriminate metal types, this feature is explored differently in this research.3 Different metal types generate positive or negative SRMMDS, suggesting the type of metal. However, the combined characteristics that compose the detected SRMMDS are fundamental for identification in this research, features such as the depth, material, posture, shape, size and soil conditions. Although previous works used databases, this research has a different approach in which a high-precision robotic arm obtains SRMMDS. Simplified, only the necessary parts of the whole SRMMDS are stored in the database using simple yet powerful mathematical relations.7,8

SRMMDS simplification. In Figure 3, θ is defined as x’y’z’, the local coordinate for SRMMDS. While the x’y’ plane parallels the MMD scanning plane, the z’-axis passes through the maximum absolute point of the SRMMDS. The plane Pθ is orthogonal to the x’y’ plane and passes through the z’-axis at an angle θ relative to the x’-axis. The intersection of plane Pθ and the SRMMDS contour generates a new curve, which is a characteristic curve known as V(r(θ)) (Figure 3) that is referenced to the new axis r(θ) and defined by the intersection of planes Pθ and x’y’.

Figure 3 demonstrates that the characteristic curves of physically symmetric targets such as anti-tank (AT) mines are the same for any angle θ, while curves for nonsymmetric targets change drastically. This analysis suggests that SRMMDS can be simplified to a set with a minimum number of characteristic curves. For symmetric cases, one characteristic curve would be enough, but this is not obvious for nonsymmetric cases. For the nonsymmetric targets (shown in Figure 2), a characteristic curve for the target’s longest length of direction presents many inflections and peaks when compared to other angles. This research defines the characteristic curve with most inflections and peaks as the main characteristic curve and its axis r(θ) as the main axis. Figure 4 shows some examples of main characteristic curves.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.

Figure 4. Examples of main characteristic curves with different peaks, intensities and sizes.

Polynomial Characterization

Characteristic curves can be represented by splines, polynomials or other mathematical relations in the form of V = f(r(θ)). As the number of inflections for the characteristic curves is limited, the authors propose polynomials in the form of Equation 1. This method has the advantage of keeping the signal characteristics and filtering part of the noisy raw data at the same time. In this work, all signals are translated with maximum peak in r = 0, making a0 the maximum absolute MMD value of the signal.

Equation 1: ƒ(Y) = a0r(θ)0 + a1r(θ)1 + a2r(θ)2 + … + anr(θ)n
Where a0, a1, a2,…, an are polynomial coefficients

In this research, the integral of the polynomials’ difference (Equation 2) is adopted as the measure of error (Err [%])—i.e., similarity—between characteristic curves, which will serve as the main criteria for discrimination.

Equation 2: Err = ∫|ƒ – g|/h*100
Where f and g are polynomials to be compared
h = max[∫|ƒ|,∫|g|]

Basic Discrimination Scheme

The basic scheme for discrimination of sensed signals can be implemented as follows:

a.
Case
Test Subject
Closest Match
Discrimination Result
R1 Metal fragment Metal fragment True negative Good: decrease FAR
R2 Metal fragment Mine False positive Still acceptable: increase FAR
R3 Mine Metal fragment False negative Not acceptable: a missed mine
R4 Mine Mine True positive Good: increase probability of detection
b.
R3’ Mine Potential mine True positive Good: increase probability of detection

Table 1.a and b. Basic discrimination cases according to Err (%). b. After the database conditioning process, case R3 becomes R3'.
All tables courtesy of authors/CISR.

This scheme can result in four possible cases, namely R1, R2, R3 or R4, as shown in Table 1(a) and illustrated in Figure 5. Cases R1 and R4 result in correct discrimination. Although R2 results in a false positive and thus increases FAR, it is still acceptable. However, case R3 finds metal fragment data as the closest match for a landmine-obtained signal, causing a false negative result (mine judged as a metal fragment), which is unacceptable in this or any other demining research.

In this research, a false negative can be overcome by flagging as potential mines all metal fragment data that can cause case R3, resulting in a new case R3’, as shown in Table 1(b). The identification of R3 and the R3’ flagging are conducted during the database building and conditioning process, as explained in the database section.

Practical Discrimination Process

Measure of difference of errors (dE). In Figure 5, the Err of some metal fragment data is close to mines, as in the R1 example. To prevent any misjudgments in a real situation, Equation 3 calculates a measure of difference of errors (dE), which is the difference between the Err of the closest metal fragment (Err(closest MF)) and the Err of the closest landmine (Err>(closest landmine)).

Equation 3: dE = Err (closest MF) – Err (closest landmine)

A threshold for dE, dEthreshold, is also defined for flagging all metal fragments in which |dE| < dEthreshold as potential mines, thereby reducing the chance that landmines are discriminated as metal fragments.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.
Figure 5. Basic discrimination cases (R1, R2, R3, R4) and target distances according to Err.

Measure of confidence (Ethreshold). Another case that can be observed in Figure 5 involves the Err of the closest target (called Eclosest) that sometimes can be too high, which indicates no matches in the database. This can mean that the data contains too much noise or the target is degraded, making it a potential risk. In this research, a safety criterion labels the test subject as a potential mine when Eclosest is greater than a given threshold, Ethreshold, to be determined by experiments. Figure 6 shows some examples of metal fragments similar to landmines.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.
Figure 6. Examples of metal fragments considered potential mines by the Eclosest and |dE| < dEthreshold criteria. Targets and corresponding depths are shown in parenthesis. Note that the International Test Operations Procedures (ITOP) conceived for an ITOP project as the metal content of larger stimulant mines shows SRMMDS very similar to the PMN2 mine, and it is also classified as a potential mine by this criteria.

Discrimination steps. The final scheme for discriminating sensed signals, while taking into account the above measures, is implemented as follows:

Database-building Experiment

In order to verify the proposed method’s validity, a database of characteristic curves (represented by polynomials) was built for multiple targets, depths and postures using a robotic manipulator. The data was taken with a metal mine-detector head at a linear speed of 50 mm/s, with a 10-mm depth step, 10-mm line step between scan lines and a signal output density of 0.2 points/mm. For the following analysis, data with weak signals (V(%) < 1%) and saturated signals (V(%) = 100%) were removed from the database.

Case
Type of Test Subject
Type of Closest Match
dE
Discrimination
R1 Metal fragment Metal fragment dE < 0, |dE| > dEthreshold True negative
R1’ Metal fragment Metal fragment flagged as “potential mine” dE < 0, |dE| ≤ dEthreshold False positive
R2 Metal fragment Mine dE > 0 False positive
R3’ Mine Metal fragment flagged as “potential mine” dE < 0 True positive
R4 Mine Mine dE > 0 True positive

Table 2. Discrimination cases: For all the above cases when EclosestEthreshold, test subject shall be considered a potential mine.

Metal detector signal conditioning. The Minelab F3 Metal Mine Detector was chosen for this experiment.15 This detector outputs signals in two independent channels (called ChA and ChB here), which are combined according to Equation 4 and detailed in endnote 16.16 ChC is used to derive characteristic curve V(r(θ)) for comparison in Equation 2.

Equation 4: ChC = ChB – ChA – median (ChB – ChA) (4)

Targets description. Figure 7 and Table 3 show target types and testing conditions. A total of 42 different targets (11 landmines and 31 metal fragments) consisting of different shapes (cubes, cylinders, spheres, tubes) and materials (aluminum, brass, chrome, stainless, steel), with depths varying from 10 mm to 400 mm, and different postures (horizontal, inclined 45° and vertical) were tested, which resulted in a total of 362 different data entries into the database. To be more applicable in an operational setting, future research efforts will increase the data library to include a range of minimum metal mines and small minefield fragments.

 
Data Number
Target Type
Dimensions
Main Composing Material
Posture
Not
Landmines
1–186 Bullets and cartridges (MFO1–MF21) 1–27 mm diameter,
27–114 mm height
Steel Horizontal
187–222 Bullets and cartridges (MFO1, MF19, MF21) 7–27 mm diameter,
27–114 mm height
Steel 45° in xz
223–254 Cube 20 mm edge Aluminum,
stainless, brass
Horizontal
255–274 Cylinder 11 mm diameter,
12.5 mm height
Aluminum,
stainless, brass
Horizontal
275–291 Tube 11 mm external diameter,
0.5 mm thickness,
12.5 mm height
Aluminum, brass Horizontal
292–301 Sphere 25.4 mm diameter Chrome Horizontal
302–305 ITOP 4.8 mm outer diameter,
0.5 mm thickness,
12.5 mm height
Aluminum Horizontal
Landmines 306–330 AT 300 mm diameter Steel Horizontal
331–335 PMN 112 mm diameter,
56 mm height
Mixture of
small alloys
Horizontal
336–340 PMN2 125 mm diameter,
65 mm height
Steel Horizontal
341–362 Other landmines
(p-40, PSM-1, MD82B, etc.)
Many variations Steel Many variations (horizontal, vertical and 45° in xz)

Table 3. Dimensions of the targets used for building the database.




Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.
Figure 7. Targets used for building the database

Database integrity and measure of confidence setting. For each given data N in the database (Table 3, N = 1 to 362), consider N as a test subject and calculate the Err (Equation 2) against all other data in the database. The cases (R1, R2, R3 and R4) described earlier are analyzed and shown (sorted for easier visualization) in Figure 8.

Figure 3. Cutting plane using as example the obtained signal of an anti-tank mine.
Figure 8. Resulting errors of closest metal fragments and mines from each data. According to the safety margins, dEthreshold and Ethreshold, different FAR can be observed.

To determine Ethreshold, several values from 0 to 100% were set, and corresponding values for false positives and true positives were observed. As Figure 9 shows, Ethreshold = 10% is the value that maximizes the difference between true positives and false positives.

Figure 9. Variation of false-positive and true-positive values according
to Ethreshold (discrete case).Figure 9. Variation of false-positive and true-positive values according
to Ethreshold (discrete case).
Figure 10. Example of polynomial interpolation for an AT target type MF21 target type. Strong relation between depths and MMD signals permit main characteristic-curves estimation by interpolating the available data in the database.Figure 10. Example of polynomial interpolation for an AT target type MF21 target type. Strong relation between depths and MMD signals permit main characteristic-curves estimation by interpolating the available data in the database.
Figure 11. Variation of false-positive and true-positive values according to Ethreshold (interpolated case).Figure 11. Variation of false-positive and true-positive values according to Ethreshold (interpolated case).
Figure 12. Trade-off of adopted safety margins and FAR. For all cases,
FAR is generated with no occurrence of false negatives due to the discrimination criteria and safety margins adopted.Figure 12. Trade-off of adopted safety margins and FAR. For all cases,
FAR is generated with no occurrence of false negatives due to the discrimination criteria and safety margins adopted.


Discrete

Interpolated

Ethreshold (%)

10

10

dEthreshold (%)

15

5

Depth margin (mm)

40

40

Metal fragments discriminated as “potential mines” according to 
Ethreshold criterion

8/14

5/14

Metal fragments discriminated as “potential mines” according to 
dEthreshold criterion

5/14

1/14

Metal fragments discriminated as landmines by closest data in database

0/14

1/14

FAR (%)

13/14 = 92%

7/14 = 50%

ITOPs discriminated as “potential mines” according to Ethreshold criterion

3/14

0/14

ITOPs discriminated as “potential mines” according to dEthreshold criterion

9/14

13/14

ITOPs discriminated as ITOP itself in vertical posture by closest data in database

1/14

1/14

Discriminated as landmine by 
closest data in database

1/14

0/14

False negatives

0/14

0/14

Time for discrimination/target (s)

< 1

< 1

Table 4. Parameters adopted and results of the proposed method.


Figure 13. FAR examples: Fragment discriminated as potential mine (left) and fragment discriminated as landmine (right). Each target's depth is shown in parenthesis. MFX and MFY are two metal fragments from the test field, of which size, shape and material are unavailable.

Expanding Database Capabilities: Data Interpolation for Different Depths

Preparing a database containing information for every depth and posture may be infeasible in reality. Fortunately, a given target’s characteristic curves basically keep the same level of concavity and mainly change in amplitude (a0) for different depths, as Figure 10 shows. For each value of r(m), MMD signals for the main characteristic curves of each depth have a quadratic relation. For example, if the input a0 is 80%, the estimated depth is around 160 mm for the AT mine and 80 mm for metal fragment 21. This strong relation between depth and signal intensities suggests that we can estimate characteristic curves from a desired depth or vice versa by interpolation (represented in red). In this work, a0 is used as input for interpolation, which generates a depth and a main characteristic curve for each target and is used for comparison in Equation 2. The data with Eclosest is then output, providing suggestions for depth, material, posture and target type.17

Repeating the analysis necessary to measure confidence setting with the interpolation method, smaller values of Err are obtained. In the new threshold, Ethreshold equals 15% (Figure 11), and R1, R2, R3 and R4 cases are set. Since no extrapolation is done in the interpolation, part of the data (each target’s deepest and shallowest data) is not used. Since depth errors are possible, depth-error margins are also considered; Figure 12 shows the analyzed trade-off.17 For interpolated cases, FAR levels are much lower when compared to the Discrete Data 10 mm case.

Figure 12 shows a FAR analysis conducted in a laboratory with the data from the database. Since potential mines were flagged with the criteria shown in the above section on discrimination, Figure 12 shows all cases in which false negatives do not occur, even if dEthreshold = 0. However, in real demining operations, dEthreshold = 0 is unacceptable, and a convenient safety margin must be set. In Figure 6, an International Test Operation Procedures (ITOP) target resembles a PMN2 mine, and it is considered a potential mine in the discrete case in which |dE| < dEthreshold criterion when dEthreshold 10%. Therefore, dEthreshold = 10% is adopted. For interpolated cases, Equation 2 identifies an ITOP target as a potential mine. While dEthreshold = 0 would be enough, a minimum of dEthreshold = 5% is adopted. Moreover, since the maximum depth estimation error of this method is 40 mm, this depth margin is adopted in real operations.17

Experimental Results

In this section, data taken in 2007 is used at a test field in Croatia.2 The Gryphon robot conducted this test. The test scanned uneven lanes of different soil properties, where several metal fragments and ITOP containing landmine surrogates were buried in random positions at depths between 1 and 14.5 cm. Among the six lanes and 38 targets per lane (180 data points in total, of which 120 were ITOP), 14 ITOP containing landmine surrogates and 14 metal fragments (bullets, rockets, etc.) were chosen to be applied as input in the proposed discrimination method. The data was chosen so that no other metal fragments were nearby, and the position was located within a standard scan area (2 sq m) to avoid cutting data. Table 4 shows the safety margins and results.

The adopted safety margins guarantee correct detection of all ITOP targets as potential mines. In the laboratory, all ITOP data (in discrete and interpolated cases) are the closest targets to metal fragment 10 (cartridge shown in Figure 7 and Table 3). In this experiment with ITOP data from the test field, six of the 14 instances for discrete cases and 12 of the 14 for interpolated cases designated metal fragment 10 as the closest target using direct search with Equation 2, which was consistent in the laboratory environment. The ITOP in the upright position in the database (not buried in the laboratory environment) and the safety margin criteria are valid for correct discrimination of data obtained with the Gryphon in soil.

A large number of the metal fragments were discriminated as potential mines in the discrete case due to the Ethreshold criterion, which indicates no similar targets exist in the database. This experiment detected eight out of 14 instances for discrete cases and five out of 14 metal fragments for interpolated cases. Due to the method’s adopted safety precautions, these results were expected. Adding similar target information to the database would result in more accurate discrimination.

Based on their proximity to some landmines, two of the 14 metal fragments were considered potential mines by dEthreshold criterion. Without available information on the test field’s metal fragment material, shape or size, they will be known as metal fragment X (MFX) and metal fragment Y (MFY). In interpolated cases, MFX was considered a potential mine for being too similar to the metal fragment 13 cartridge (Figure 7 and Table 3) and was also considered a potential mine for being too similar to the PMN2 landmine (Figure 13). MFY was identified as a landmine by direct search, in which a PMN2 was identified as the closest data match (Figure 13).

The better performance of the interpolated method generates lower FAR levels. Time is another great advantage of using this method; it takes one second per target, which is faster than the false-alarm reduction method endnote 9 references, which takes more than 96 seconds per target.9

Conclusions

  Discrete Interpolated
Ethreshold (%) 10 10
dEthreshold (%) 15 5
Depth margin (mm) 40 40
Metal fragments discriminated as “potential mines” according to
Ethreshold criterion
8/14 5/14
Metal fragments discriminated as “potential mines” according to
dEthreshold criterion
5/14 1/14
Metal fragments discriminated as landmines by closest data in database 0/14 1/14
FAR (%) 13/14 = 92% 7/14 = 50%
ITOPs discriminated as “potential mines” according to Ethreshold criterion 3/14 0/14
ITOPs discriminated as “potential mines” according to dEthreshold criterion 9/14 13/14
ITOPs discriminated as ITOP itself in vertical posture by closest data in database 1/14 1/14
Discriminated as landmine by
closest data in database
1/14 0/14
False negatives 0/14 0/14
Time for discrimination/target (s) < 1 < 1
Table 4. Parameters adopted and results of the proposed method.

 

The above tests of this new methodology for the discrimination of landmines and metal fragments using commercially available MMDs and a prebuilt library demonstrate that this methodology can lead to effective signal characterization and real-time discrimination. Moreover, the methodology to interpolate discrete data into the database according to its depth makes the evaluation of data in arbitrary depths possible. False positives, which increase FAR, depend on the adopted error-margin criteria. After extensive laboratory tests, thresholds of Ethreshold (%) = 15% and dEthreshold (%) = 5% were selected, which reduces the FAR to about 50%.

Results from the data analysis obtained in a Croatian test field in 2007 showed the robustness, validity and potential of the proposed method for practical applications. This technology could also potentially help detect unexploded ordnance (UXO) as well. However, additional testing with real UXO and mines, especially low-metal mines, will be needed if that application is pursued. Further tests in real minefields are in journal as the next step in this work. This includes tests scheduled for 2014 in Angola that will investigate more types of landmines and metal fragments, as well as other important factors such as soil and climate. c

JSPS KAKENHI Grant Number 25303012 supported this work.

Biographies

Inna CruzAlex M. Kaneko received a Bachelor of Engineering in mechatronics engineering from the University of Sao Paulo and a Master of Engineering in mechanical and aerospace engineering from the Tokyo Institute of Technology, where he is a doctoral candidate. His research activities include development of demining robots.


Daniel ErikssonEdwardo F. Fukushima is an associate professor at the Department of Mechanical and Aerospace Engineering of the Tokyo Institute of Technology. His research activities include the development of demining robots, design of controllers for intelligent robots and journal of new brushless motors and drives.


Daniel ErikssonGen Endo is an assistant professor at the Department of Mechanical and Aerospace Engineering of the Tokyo Institute of Technology. His research interests include mechanical design and intelligent control of mobile robots, especially in legged robots, leg-wheel hybrid robots, assistive mobile robots, educational robots and demining robots.



Contact Information

Alex Masuo Kaneko
Ph.D. Candidate
Department of Mechanical and Aerospace Engineering
Graduate School of Science and Engineering
Tokyo Institute of Technology
2-12-1 Ookayama, Meguro-ku
Tokyo 152-8552 / Japan
Tel/Fax: +81 3 5734 2648
Email: junjyouna.amk@gmail.com

Edwardo F. Fukushima, Dr. Eng.
Associate Professor
Department of Mechanical and Aerospace Engineering
Graduate School of Science and Engineering
Tel/Fax: +81 3 5734 3175
Email: fukusima@mes.titech.ac.jp

Gen Endo, Dr. Eng.
Assistant Professor
Department of Mechanical and Aerospace Engineering
Graduate School of Science and Engineering
Tel/Fax: +81 3 5734 2774
Email: gendo@mes.titech.ac.jp

 

Endnotes

  1. Debenest, P., E.F. Fukushima, Y. Tojo, and S. Hirose. “A New Approach to Humanitarian Demining - Part 1: Mobile Platform for Operation on Unstructured Terrain.” Autonomous Robots 18 (2005) 303–321. http://download.springer.com/static/pdf/145/art%253A10.1007%252Fs10514-005-6842 9.pdf?auth66=1391002924_76a097213520b187ec11d251a528f469&ext=.pdf.
  2. Pavkovic, N., J. Ishikawa, K. Furuta, K. Takahashi, M. Gaal, and D. Guelle. “Test and Evaluation of Japanese GPR-EMI Dual Sensor Systems at Benkovac Test Site in Croatia.” Anti-personnel Landmine Detection for Humanitarian Deming, edited by Katsuhisa Furuta and Jun Ishikawa, 63–81. London: Springer, 2009.
  3. Minelab. “The Minelab Eureka Gold.” Instruction Manual.
  4. Ho, K.C., L.M. Collins, L.G. Huttel and P.D. Gader. “Discrimination Mode Processing for EMI and GPR Sensors for Hand-Held Land Mine Detection.” IEEE Transactions on Geoscience and Remote Sensing 42, no. 1 (January 2004) 249–63. http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=1262601&tag=1.
  5. Riggs, L.S., J.E. Mooney, and D.E. Lawrence. “Identification of Metallic Mine like Objects Using Low Frequency Magnetic Fields.” IEEE Transactions on Geoscience and Remote Sensing 39, no. 1 (January 2001) 249–63. http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=898665&userType=inst.
  6. Gao, P., L. Collins, P.M. Garber, N. Geng, and L. Carin. “Classification of Landmine-Like Metal Targets Using Wideband Electromagnetic Induction.” IEEE Transactions on Geoscience Remote Sensing 38, no. 3 (May 2000) 1,352–61. http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=843029&userType=inst.
  7. Kruger, H., and H. Ewald. “Handheld metal detector with online visualization and classification for the humanitarian mine clearance.” Sensors, 2008 IEEE (2008): 415–18.
  8. Freese, M.A. “Improved Landmine Discrimination With an Off-the-Shelf Metal Detector.” The Journal of Mine Action 12, no. 1 (2008): 93–99.
  9. Narita, T., E.F. Fukushima, and S. Hirose. “Journal of False Alarm Reduction Method of Metal Detector with Robotic Arm for Humanitarian Demining.” Paper presented at the JSME Robotics and Mechatronics Conference, Okayama, Japan, 26–28 May 2011.
  10. Stanley, R.J., K.C. Ho, P. Gader, J.N. Wilson, and J. Devaney. “Land Mine and Clutter Object Discrimination Using Wavelet and Time Domain Spatially from Metal Detectors and Their Fusion with GPR Features for Hand-Held Units.” Circuits Systems Signal Processing 26, no. 2 (2007) 165–91. http://download.springer.com/static/pdf/505/art%253A10.1007%252Fs00034-005-1205-5.pdf?auth66=1391004470_0b481f85bacdde3013e24656c5e75ed6&ext=.pdf.
  11. Collins, L., P. Gao, D. Schofield, J.P. Moulton, L.C. Makowsky, D.M. Reidy, and R.C. Weaver. “A Statistical Approach to Landmine Detection Using Broad Band Electromagnetic Induction Data.” IEEE Transactions on Geoscience and Remote Sensing 40, no. 4 (April 2002) 950–962. http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=01006387.
  12. Doheny, R.C., Sean Burke, Roger Cresci, Peter Ngan, and Richard Walls. “Handheld Standoff Mine Detection System (HSTAMIDS) Field Evaluation in Thailand.” U.S. Army Research journal and Engineering Command, 2005.
  13. Sato M., J. Fujiwara, X. Feng, Z. Zhou, and T. Kobayashi. “Evaluation of a Hand-held GPR MD sensor system (ALIS).” Proc. SPIE Vol. 5794: Detection and Remediation Technologies for Mines and Minelike Targets X (2005): 1,000–07.
  14. Debenest, Paulo, Marc Freese, Edwardo F. Fukushima, Toshiaki Matsuzawa, and Shigeo Hirose. “Lessons Learned from Field Tests in Croatia and Cambodia.” The Journal of Mine Action 11, no. 2 (2008): 103–09.
  15. Minelab Eletronics. “F3 Metal Mine Detector - Instructors Notes and Syllabus.” (2006).
  16. Kaneko, A.M., and E.F. Fukushima. “journal of an Automatic Landmine Detection and Marking System for the Demining Robot Gryphon.” Journal of Advanced Computational Intelligence and Intelligent Informatics 15, no. 6 (2011): 737–43.
  17. Kaneko, A.M., E.F. Fukushima, and G. Endo. “Landmine Buried Depth Estimation by Curve Characterization of Metal Mine Detector Signals.” Paper presented at the IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS), Tokyo, Japan, 3–7 November 2013.

 

TOP OF PAGE