This article introduces Japanese activities regarding a project, “Research and Development of Sensing Technology, Access and Control Technology to Support Humanitarian Demining of AP Mines.” This project, which includes the research of six teams from academia and industry, has been funded by the Japan Science and Technology Agency (JST) under the auspices of the Ministry of Education, Culture, Sports, Science and Technology (MEXT). The developed systems are equipped with both ground-penetrating radar and a metal detector, and they are designed to make no explicit alarm and to leave decision-making of detection using subsurface images to the operators. To evaluate these kinds of systems, a series of trials was conducted in Japan from 8 February to 11 March 2005.
At current clearance speed, it will take more than 100 years to remove all the landmines that remain in the world.1 Consequently, Japan is developing more efficient and safer humanitarian-demining technologies. This article introduces Japanese robotic sensor systems that provide deminers with clear subsurface images via ground-penetrating radar in combination with metal detectors (GPR+MD).
Experiment Overview: Background
To reconstruct clear images, highly accurate sensor-positioning systems, as well as sensing technology itself, are indispensable because one of the most important pieces of information for signal processing is sensor position, where the sensor acquires a series of data for GPR+MD.
There are many kinds of anti-personnel landmines, which can be laid by humans or scattered by airplanes, and mined areas are not limited to plains but also marshes, canals, steep hillsides, seashores, deserts, mountains and forests. For such rough terrain, robotic systems must have sensor heads that can scan the ground as closely as possible but never touch it as well-trained deminers do. Metal detectors, which are a kind of an electromagnetic induction (EMI) sensor, have the possible detection distance of about 15 centimeters for minimum-metal landmines. For these metal detectors, it is a challenge for sensor systems to access minefields and manipulate the sensor head in severe environments in order to stay as close to the ground as possible. Thus, Japanese advanced robotics and sensor engineering have been fused to create novel detectors.
Japan started preparation for this kind of research and development in March 1997, when the Tokyo Conference on Anti-personnel Landmines was held. At this conference, participants undertook a comprehensive discussion to strengthen international efforts toward addressing the problems of AP landmines, especially landmine clearance by the United Nations and other organizations; development of new technology for mine detection and removal; and assistance to victims. In December 1997, Keizo Obuchi, then Minister for Foreign Affairs of Japan, signed the Ottawa Convention,2 and the ultimate goal of zero victims was proposed. Since August 2002, the Japanese have undertaken preparations to start humanitarian-demining R&D.3
Japanese R&D of Anti-personnel Landmine Detection System
With strong expectations from the world community for Japanese contributions in this area, the Ministry of Education, Culture, Sports, Science and Technology established the Committee of Experts on Humanitarian Demining Technology in January 2002, believing in the importance of tackling the technological development of AP landmine detection using advanced Japanese technology. The Committee’s findings were presented to MEXT in the report, “Promoting R&D for Humanitarian Demining Technology.”4 Based on this report, the Japan Science and Technology Agency announced a call for proposals for R&D projects in humanitarian-demining technology. Out of the 82 proposals, 12 projects were selected, and an R&D project named “Research and Development of Sensing Technology, Access and Control Technology to Support Humanitarian Demining of Anti-personnel Mines” started in October 2002.
The JST project is essentially divided into a short-term R&D project and a medium-term one. Because of the urgent need for this technology, the short-term R&D project is expected to have prototypes in field trials within three years. The JST medium-term R&D project is on a five-year schedule. The goal is to develop sensing technologies that can detect the explosive itself, in the range of about 30 to 100 grams.
Short-term R&D project. The objectives of the short-term R&D project is to develop sensing technology that can safely and efficiently detect AP landmines based on the physical differences between landmines and soils, and to develop access devices and manipulation technology that carry sensors into minefields and allow them to scan the ground precisely. More specifically, the goal is to develop vehicle-mounted GPR+MD dual-sensor systems that make no explicit alarm and provide operators with clear subsurface images. This means that the decision to determine whether or not a shadow in the image is a real AP landmine is entirely left to the operator, similar to how medical doctors can find cancer by reading CT images. This feature discriminates the systems from conventional GPR+MD dual sensors that are based on alarm tones.
In the short-term project, four sensors and three robotic vehicles have been developed. One of those is the Mine Hunter Vehicle (Figure 1). The vehicle itself and the manipulator have been developed by a research team of Professor Kenzo Nonami’s at Chiba University.5 The MHV can interchangeably mount two GPR sensors in addition to a commercial, off-the-shelf metal detector.
One sensor is a stepped-frequency GPR developed by Professor Motoyuki Sato’s team at Tohoku University,6 hereinafter referred to as MHV #1 (Figure 2). Stepped-frequency radar determines distance to a target by constructing a synthetic range profile, which is a time domain approximation derived from the frequency response of a combination of stepped-frequency signals via inverse fast Fourier transform (IFFT). The major advantage of the stepped-frequency method is that the spectrum bandwidth can be easily tuned to fit an optimum value according to environmental conditions such as soil moisture.
The other sensor is an impulse GPR developed by Professor Ikuo Arai’s project at the University of Electro-Communications,7 hereinafter referred to as MHV #2. This kind of GPR operates by transmitting a very narrow pulse (< 1 nanosecond) of electromagnetic wave, the advantage of which is that the measurement time required to generate one range profile is very short. After the GPR scans to acquire a range profile for every interval of several centimeters,8 GPR tomography gives subsurface horizontal slices as shown in Figure 3a, and further calculation provides operators with three-dimensional images (Figure 3b).
Professor Toshio Fukuda’s group at Nagoya University developed a dual sensor with built-in stepped-frequency GPR+MD9 (Figure 4). The sensor system scans the ground, being carried by a low-reaction-force manipulation frame that has four balloons on the legs to softly land it on minefields. The manipulation frame is attached to the top of a boom of a crane vehicle developed by Mr. Tomohiro Ikegami’s group at TADANO Ltd. The vehicle has a 20-meter reach for a 200-kilogram payload with a positioning accuracy of 15 centimeters. These elements have been integrated into the Advanced Mine Sweeper (AMS) (Figure 5), which can adapt to various geographical environments.10
Professor Shigeo Hirose’s team at the Tokyo Institute of Technology developed the Gryphon buggy system (Figure 6), which can be remotely controlled to access minefields.11 The manipulator mounted on the buggy has been designed to cancel reaction force induced by sensor scanning.12 The sensor is a GPR+MD dual sensor named the Advanced Landmine Imaging System (ALIS), and it can also be used as a handheld detector.13 ALIS was developed by Professor Sato’s team and underwent a field trial in Afghanistan in December 2004.
Medium-term R&D project. Professor Hideo Itozaki’s group of Osaka University is developing a nuclear quadrupole-resonance detector.14 In the analysis, a radio-frequency electromagnetic wave is first emitted and excites nuclear spin of 14N in explosives. Then a magnetic wave detector, such as an induction coil, detects subsequent NQR signals from the 14N if any intended target exists, and the resonance frequency of the signal is unique for each explosive material. Thus explosives can be identified.
Two research teams on the project are trying to develop detectors based on the neutron analysis identifying explosives through backscattering of neutrons and detection of specific energy gamma rays from capture on hydrogen and nitrogen atoms of explosives (Figure 7). Professor Kiyoshi Yoshikawa’s group from Kyoto University has prototyped an extremely compact neutron source based on an inertial-electrostatic confinement fusion device 20 centimeters in diameter.15 Professor Tetsuo Iguchi’s group of Nagoya University has prototyped another neutron source, which is an improved Cockcroft-Walton-type accelerator neutron source using a deuterium-deuterium (DD) fusion reaction. They have also developed a prototype of a multi-Compton γ camera, which estimates the incoming direction of 10.8MeV γ-rays produced from the nitrogen of the explosive (Figure 8).16
The medium-term R&D project is expected to have prototypes in field trials within five years, namely in 2007, in combination with one of the prototypes of MHV, AMS or Gryphon.
To evaluate the short-term R&D prototypes, a series of tests was conducted from 8 February to 11 March 2005 in Sakaide City, Japan. Seven test lanes were constructed using more than 200 landmine surrogates (Figure 9, Figure 10). Since operators’ pre-knowledge of the locations of buried targets significantly influences the detection results for such systems that make no explicit alarm, lanes 1 to 6 are designed to be used for blind tests.
Test lanes and landmine surrogates. In constructing test lanes, all the original soil was removed from a width of 2 meters to a depth of 0.5 meters in the vertical section, and the lanes were filled with homogeneous and non-mineralized (cooperative) soil. The actual width of test lanes is 1 meter, and mine surrogates were buried shallower than or equal to a depth of 0.3 meters. The features of each lane are as follows:
- Lanes 1, 2 and 3 are 15 meters long with a flat surface.
- Lane 4 is 20 meters long, with 15 bumps in the surface, each with a height of 10 centimeters and a diameter of 60 centimeters, and small stones are mixed to make the soil heterogeneous.
- Lane 5 simulates minefields in post-clearance inspection after mechanical demining, with the soil stirred and not packed.
- Lane 6 is wet, with 10 liters of water per square meter sprinkled one hour before the test starts.
Figure 11 shows four kinds of landmine surrogates used in the test. The M1419 and PMN220 contain a metal part—an 18-millimeter21 vertical carbon steel pin with a diameter of 3 millimeters—and the Type 7222 has a 4-millimeter vertical carbon steel pin with a diameter of 4 millimeters. The Type72-S23 mine is made by modifying a product of Amtech Aeronautical Limited24 and has exactly the same metal part as the International Test Operations Procedure standard I0, a 12.7-millimeter vertical aluminum tube (Figure 12). Silicone rubber was substituted for explosives in all the surrogates.
Experimental design. Through the tests, influences of various factors on probability of detection should be evaluated. Namely, in Experiment 1, target types, target depth, soil conditions and target angles were chosen as factors to be tested (Figure 13). There are two or four levels for each factor as described in Table 1.
Table 1: Factors A to D and the levels for Experiment 1.
According to the soil conditions, for example, targets (landmine surrogates) that are classified into “flat,” “wet,” “stirred” and “rough” are respectively buried in lanes 1–3, 6, 5 and 4 at a specified depth and angle as defined in Figure 13. Experiment 2 was designed to mainly evaluate the minimum discrimination distance. Two levels were chosen for the factor “distance to adjacent target” as described in Table 2. One level consists of pairs of targets in a distance of 15 centimeters and the other level consists of independent targets, the separation of which shall be at least 50 centimeters.
Table 2: Factors A to C and the levels for Experiment 2.
Due to the limitation of time for the trial and the number of targets, it is impossible to test all the combinations of levels in Tables 1 and 2. To impartially collect unbiased data for statistical analysis under this limitation, orthogonal experimental designs based on L16 (215) and L8 (27) orthogonal arrays were respectively used for Experiments 1 and 2. Assigning the columns of the array to each factor as specified in Tables 1 and 2 derives a reduced set of combinations, the results of which are summarized in Tables 3 and 4. For example, the number of experimental runs can be reduced from 128 (4×4×4×2) to 16 in Experiment 1.
Table 3: Design result for Experiment 1.
Table 4: Design result for Experiment 2.
According to Tables 3 and 4, all the targets were buried at random locations in the specified lanes and were left for more than one month before the test began. Testees can submit all the impartial data needed for statistical analysis by reporting detection results from lanes 1 through 6. In the trial, at least two testees from every device took the test in all 6 lanes.
Benchmarking. To compare performance of the GPR+EMI dual sensors with that of existing metal detectors, a benchmarking trial was conducted. Namely, a tester who knew the exact positions of targets checked if any metal-detector response occurred just above every buried target. The result of this test shows the best performance of the metal detectors used.
Test procedures. Testees took blind tests for each lane following the procedures as described below:
- Before the test starts, the tester records temperature, relative humidity and volumetric water content that is measured by time domain reflectometry (TDR).25
- The testee does close-in detection work using a sensor system cooperatively with vehicle operators.
- After the work finishes, the tester records temperature, relative humidity and volumetric water content measured by TDR.
- The testee reports the following data for every detected anomaly:
- Coordinates of the detected target
- Depth of the detected target
- Confidence rating defined in Table 5 and the final decision whether or not to declare the anomaly as a landmine surrogate
- True positive: The testee declared it as a target and this is true.
- False positive: The testee declared it as a target and this is not true. This is a false alarm.
- True negative: The testee declared it as a fragment, clutter or noise and this is true.
- False negative: The testee declared it as a fragment, clutter or noise and this is not true. This is missing a target.
Table 5: Definition of confidence rating.
Completing the tests from lanes 1 through 6 means that the testee finished all 24 experimental runs of Experiments 1 and 2 described in Tables 3 and 4.
The most important thing is to practically use these technologies to improve landmine-detection efficiency and reduce minefields. To do so, the mine-detection systems must be robust, simple and highly cost-effective. The Japanese domestic trial is the first step.
Test and Evaluation Results
The following is the data analysis and evaluation of test results for anti-personnel landmine detection systems using ground-penetrating radar mounted on robotic vehicles for humanitarian demining.17,18 The test results showed that combining GPR with metal detectors can improve probability of detection for targets around a depth of 20 centimeters, where it is difficult to detect the targets by using only a metal detector. It has also been learned that positioning control must be improved in scanning the ground with a sensor head, which is key to making the best of use of metal detectors mounted on vehicles. Lessons learned have been reflected in further improvement of the prototypes. In the following sections data analysis, methods and evaluation results are described.
Data analysis. According to the experimental design proposed above,
data from eight testees (two each from every system) have been
acquired. The comprehensive results of probability of detection (PD)
are shown in Tables 6 and 7 and were acquired through Experiments 1
and 2. The systems named are anonymous and described as Device 1, 2,
3 and 4. A benchmarking result is also shown in the tables. This
section discusses how the data are analyzed.
Table 6: PD of eight testees of Experiment 1. Highlighted data of four testees are analyzed as shown in Figure 21.
Table 7: PD of eight testees of Experiment 2. Highlighted data of four testees are analyzed as shown in Figure 22.
Analysis of variance (ANOVA). ANOVA tests are necessary if there are significant differences of PD between levels for each factor.28 This is useful to check if experiments are well designed to discuss influences of the factors on PD and to see how the factors interfere in PD. Some levels such as a target depth of 30 centimeters have been set to be very difficult in comparison with the sensor specifications because an objective of the test is to make the limitations of the sensor systems clear.
In the following part of this section, an example is given for an ANOVA of Experiment 2, assuming that an experimental result in Table 8 is acquired from a system with no repetition. First the mean of the results is calculated as:
and the main effect for each level of the factors A, B and C is derived as follows:
Table 8: Notion of detection of probability for ANOVA example.
Next, error effects for are calculated as:
Now, for example, a linear model for the probability of detection can be defined as:
For the ANOVA, four means of squares (variances) are calculated as follows:
where ,,, and are the degrees of freedom of factors and error.
By comparing the variances due to levels of each factor (i.e.,
with the variance due to measurement error  using F-test), the
significance of the differences between levels is tested. In this
test, the null hypothesis is that the main effects of levels for a
factor are all equal (i.e., there is no difference in influences of
levels for the factor on PD). The computed F statistic in Table 9
follows an F distribution with corresponding degrees of freedom
under the assumption that variances of PD have homogeneity.29
Therefore, the significance of F can be determined in the usual way
by using the table of F. If the computed value of F is larger than
the tabled value, the null hypothesis is rejected. This means that
at least one pair of main effects is significantly different.
Table 9: Analysis of variance (ANOVA).
The 95-percent confidence limit of each main effect is experimentally derived by using , the mean of squares due to error. For example, the 95-percent confidence interval of is given by:
where is the total number of experiments (the number of experimental runs multiplied by repetitions), and is the quantile of the t-distribution for probability 95 percent with degrees of freedom.
Receiver operating characteristic curve. It has been 30 years since radiographic applications of ROC curves were reported30 and it is well-known that analysis based on ROC curves is suitable for subjective evaluation of imaging equipment. In the test and evaluation here, ROC curves were also used to evaluate sensor effectiveness in terms of both PD and false-alarm rate.
As described above, detection results reported by testees are classified into four categories: true positive, false positive, true negative and false negative. However, the classification based on a testee’s discrimination threshold is a one-sided view, and the number of true positives and the number of false positives change as the threshold is varied. An ROC curve shows us the relationship between the true positive and false positive for a variety of different thresholds, thus helping the determination of an optimal threshold as well as the comparison of sensor performance.
Figure 14: Normalized histogram of signal and noise.
Figure 15: Example of ROC curves.
To plot an ROC curve, two histograms, which are measured on an interval scale in the confidence rating reported by the testee, are needed. One is from signals of intended targets that consist of true positives and false negatives, and the other is from signals of fragments, clutters or noise (i.e., true negatives and false positives). According to the histograms, the ratio of true positive (i.e., probability of detection) is plotted as a function of the ratio of false positive at every confidence rating (threshold). As shown in Figure 14, if a sensor functions well, a histogram of targets (solid line) is distributed apart from that of noise (dotted line), and the resulting ROC curve climbs rapidly toward the upper left-hand corner of the graph as shown by the solid line in Figure 15. On the other hand, if another sensor gives a histogram of targets that is distributed closer to that of noise, the resulting ROC curve gets closer to a diagonal line as shown by the dashed line in Figure 15. This means the discriminating power decreases. Once ROC curves are obtained, there are many methods to test the difference between ROC curves.31
In the experiment, the number of true positives is controlled, but the number of false positives depends on how many false alarms are reported by the testee. Therefore, all the histograms discussed here are normalized by dividing frequencies by the total number of the population.
Figure 16 shows the ground truth of the lane 2, and Figures 17 and 18 shows subsurface images from a sensor system. In this case, it has been shown that a metal detector can clearly image seven pairs of Type 72 surrogates buried flush (Figure 17), and that a GPR sensor can display seven PMN2 surrogates at a depth of 20 centimeters (8 inches) (Figure 18), where the metal detector was not able to get any signal. Based on these kinds of images, testees have derived their detection results, and this section discusses the experimental results.
Probability of detection. The number of testees is eight, the breakdown of which is two from MHV with a step-frequency GPR+MD (MHV #1), two from MHV with a pulse GPR+MD (MHV #2), two from the Advanced Mine Sweeper with a step-frequency GPR+MD, and two from Gryphon with a pulse GPR+MD. The eight sets of data were analyzed by ANOVA to see the effects of factors. Note that the order of the systems is not consistent with devices 1–4 to keep anonymity.
Tables 10 and 11 show ANOVA results for Experiments 1 and 2,
respectively, and Figures 19 and 20 show plots of factor effects
(i.e., main effects added to the mean μ with 95-percent confidence
intervals derived in the same way as Equation 11). In Tables 10 and
11, factors, the null hypothesis of which has been rejected at the
level of significance of 0.05/0.01, are indicated by * (0.05) /**
(0.01). For those factors, there have been significant differences
in PD between the levels, and it can be said that it is meaningful
to discuss how those factors influence PD and that the test lanes
were well-designed to evaluate the sensor systems. It has been shown
that there is a strong dependence of PD on target depth and that the
developed systems still have problems for rough and uneven ground
surface (Figures 19 and 20). Regarding factor A of Experiment 2,
distance to adjacent target, the ANOVA showed that there was no
significant difference in PD between a pair of Type 72 S surrogates
at a 15-centimeter distance and the other independent
Type 72 S surrogates.
Table 10: Eight testees’ result of ANOVA of Experiment 1.
Table 11: Eight testees’ result of ANOVA of Experiment 2.
Figure 21: Averages of PD for Experiment 1. Testees 7, 2, 3 and 6 were chosen from each device.
(click on thumbnail to see larger image in new window)
Figure 22: Averages of PD for Experiment 2. Testees 7, 2, 3 and 6 were chosen from each device.
(click on thumbnail to see larger image in new window)
Averages of PD of four testees, that is, one each from every system, are plotted in Figures 21 and 22, compared with the benchmarking result using only a metal detector. Confidence limits can be calculated in such the way K. M. Simonson discusses in the Sandia Report32,33 as the number of population for each level is derived from Tables 10 and 11 above. These results showed that the PD for targets deeper than 10 centimeters can be improved by combining GPR with MD. On the other hand, as also shown in Figures 21 and 22, some of the GPR+MD results in shallow levels were worse than those of metal detectors. This is because sensor height above the ground, which is controlled by manipulators, is higher than that of manual scanning of metal detectors, and this is considered to be improved by modifying the manipulation algorithm of a robotic part.
Lessons learned. Through the test and evaluation process, many lessons have been learned, some of which are listed below:
- The provided calibration area should have contained landmine surrogates for all levels of factors. Coaching a typical image for each level would much improve the detection rate.
- In some cases (for example, Testee 7) high PDs have been accompanied by high false-alarm rates around 30 times/square meter,34 and it was also proven that confirming the source of false alarms for GPR is much more difficult than those of metal detectors (i.e., metal fragments). Therefore, another performance index to penalize these GPR false alarms will be needed.
- PD in deep levels of 20–30 centimeters can be improved by combining GPR with MDs.
Evaluation of FAR. As described above, ROC curves are useful to see
the qualification of sensors, taking into account tradeoff between
PD and false alarm rate. Table 12 shows the FAR of eight testees for
each of the six lanes in the experiment.
Table 12: False-alarm rate (1/square meter) of eight testees for each lane.
Figures 23a through 23d show typical ROC curves of testees 7 and 3 for lanes 2 and 4. Lane 2 has 21 targets buried as shown in Figure 16 (see above), and lane 4 with rough ground surface has 77 targets. A horizontal axis of each plot shows the normalized FAR, and the number of false alarms can be derived by FAR multiplied by the total number of negatives that is shown in each plot. In the case of Figure 23a, 65 percent of targets were detected with 100-percent confidence, but the other targets got mixed in 525 negatives. In Figure 23b, 95 percent of the targets were detected with 100-percent confidence. Figure 23c for lane 4 showed that testee 7 could not discriminate the targets from 738 negatives although the PD was 77 percent. On the other hand, as shown in Figure 23d, testee 3, the PD of which was 50 percent, detected 40 percent of targets with 100-percent confidence. These kinds of data have been used to optimize operator’s decision threshold and sensor sensitivities, and to improve the sensor performance.
|Figure 23a: ROC curve for lane 2 (testee 7). The total number of negatives (fragments, clutters or noise) is shown.||Figure 23b: ROC curve for lane 2 (testee 3). The total number of negatives (fragments, clutters or noise) is shown.|
|Figure 23c: ROC curve for lane 4 (testee 7). The total number of negatives (fragments, clutters or noise) is shown.||Figure 23d: ROC curve for lane 4 (testee 3). The total number of negatives (fragments, clutters or noise) is shown.|
Through the test and evaluation, many lessons have been learned, and
these results were fed back to the testees for further improvement.
The next step of the project is field trials in some mine-affected
countries to confirm the improvements and to evaluate robustness and
The authors would like to thank all the project members, especially the principal partners for the trial: Tohoku University, Chiba University, Tokyo Institute of Technology, University of Electro-Communications, Nagoya University, Kyoto University, Osaka University, TADANO Ltd., Mitsui Engineering & Shipbuilding Co., Fuji Heavy Industries Ltd., TAU GIKEN Co. Ltd. and Tokyo Gas Co.
Jun Ishikawa received bachelor’s and master’s degrees in control engineering from the Tokyo Institute of Technology in 1989 and 1991, respectively. He joined NEC Corporation in 1991 as a Research Engineer of mechatronics and computer peripherals. He has been a research manager of the R&D Office for Supporting Anti-personnel Mine Detection and Removal Activities for the Japan Science and Technology Agency since 2002.
Mitsuru Kiyota received a Bachelor of Engineering from KEIO University in 1971. He joined Mitsubishi Electric Corporation in 1971. He was engaged in the development of defense equipments. He was transferred to the Japan Science and Technology Agency in 2002 and is now a Research Manager of the R&D Office for Supporting Anti-personnel Mine Detection and Removal Activities.
Katsuhisa Furuta received his Bachelor and Master of Science and doctorate degree in engineering from the Tokyo Institute of Technology in 1962, 1964 and 1967, respectively. He is currently a professor, a member of the Board of Trustees and Director of the 21st Century Center of Excellence Project at Tokyo Denki University and the supervisor of the MEXT/JST Project of Humanitarian Anti-personnel Mine Detection and Removal Activities.
- “Church World Service: Emergency Response Program. Revised Landmine Appeal.” Center for International Disaster Information. http://iys.cidi.org/humanitarian/hsr/01a/ixl15.html. Accessed May 1, 2006.
- Convention on the Prohibition of the Use, Stockpiling, Production and Transfer of Anti-Personnel Mines and on Their Destruction, Oslo, Norway. 18 Sept. 1997; http://www.un.org/Depts/mine/UNDocs/ban_trty.htm. Accessed 26 April 2006. The document was opened for signature in Ottawa, Canada, 3 Dec. 1997, and thus is commonly known as the Ottawa Convention.
- K. Furuta, “Japan Promises New Technologies to Clear Anti-personnel Landmines,” Look Japan, vol. 48, No. 562, January 2003.
- Committee of Experts on Humanitarian Demining Technology, “Promoting R&D for Humanitarian Demining Technology,” Ministry of Education, Culture, Sports, Science and Technology of Japan (MEXT), 2002. http://www.mext.go.jp/english/news/2002/05/020601.htm. Accessed April 12, 2006.
- K. Nonami and H. Aoyama, “Research and Development of Mine Hunter Vehicle for Humanitarian Demining,” Proceedings of the IARP International Workshop on Robotics and Mechanical Assistance in Humanitarian Demining (HUDEM2005), pp. 76–81, 2005.
- M. Sato, Y. Hamada, X. Feng, F. Kong, Z. Zeng, G. Fang, “GPR using an Array Antenna for Landmine Detection,” Near Surface Geophysics, Vol. 2, pp. 3–9, February, 2004.
- S. M. Shrestha and I. Arai, “High-Resolution Image Reconstruction by GPR using MUSIC and SAR Processing Method for Landmine Detection,” Proceedings of the 2003 IEEE International Geoscience and Remote Sensing Symposium (IGRASS2003), pp. 505–508, 2003.
- There are 2.54 centimeters in one inch.
- Y. Hasegawa, K. Yokoe, Y. Kawai and T. Fukuda, “GPR-based Adaptive Sensing—GPR Manipulation According to Terrain Configurations,” Proceedings of the 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS2004), pp. 3021–3026, 2004.
- T. Fukuda, et al., “Environment-Adaptive Anti-personnel Mine Detection System—Advanced Mine Sweeper,” Proceedings of the HUDEM2005, pp. 33–38, 2005.
- E. F. Fukushima, et al., “Teleoperated Buggy Vehicle and Weight Balanced Arm for Mechanization of Mine Detection and Clearance Tasks,” Proceedings of the HUDEM2005, pp. 58–63, 2005.
- Y. Tojo, P. Debenest, E. F. Fukushima and S. Hirose, “Robotic System for Humanitarian Demining: Development of Weight-Compensated Pantograph Manipulator,” Proceedings of the 2004 IEEE International Conference on Robotics and Automation (ICRA2004), pp. 2025–2030, 2004.
- M. Sato, J. Fujiwara, X. Feng, Z. Zhou and T. Kobayashi, “Development of a Handheld GPR MD Sensor System (ALIS),” Proceedings of SPIE Vol. 5794, Detection and Remediation Technologies for Mines and Mine-like Targets X, pp. 1000–1007, 2005.
- M. Tachiki, et al., “Remote Detection of Nitrogenated Substances by Nuclear Quadrupole Resonance,” Proceedings of the HUDEM2005, pp. 107–109, 2005.
- K. Yoshikawa, et al., “Research and Development of Humanitarian Landmine Detection System by a Compact DisCharge-Type Fusion Neutron Source,” Proceedings of the HUDEM2005, pp. 114–117, 2005.
- T. Iguchi, et al., “Development of Compact Compton Gamma Camera for Anti-personnel Landmine Detection with Neutron Induced Prompt Gamma-ray Imaging,” Proceedings of the HUDEM2005, pp. 110–113, 2005.
- J. Ishikawa, M. Kiyota, and K. Furuta, “Experimental Design for Test and Evaluation of Anti-personnel Landmine Detection Based on Vehicle-mounted GPR Systems,” Proceedings of SPIE Vol. 5794, Detection and Remediation Technologies for Mines and Minelike Targets X, 2005.
- J. Ishikawa, M. Kiyota, and K. Furuta, “Evaluation of Test Results of GPR-based Anti-personnel Landmine Detection Systems Mounted on Robotic Vehicles,” Proceedings of the HUDEM2005, pp. 39–44, 2005.
- The M14 is a small, plastic blast mine. It is hard to detect with metal detectors because it has little metallic content. For more information visit, http://www.hrw.org/reports/1997/gen2/General2-06.htm. Accessed May 10, 2006.
- The PMN2 is a Russian mine that contains a large amount of explosive. The way it is designed makes it nearly impossible to neutralize. For more information visit http://www.fourmilab.ch/minerats/figures/mines.html. Accessed May 15, 2006.
- One millimeter is equivalent to 0.04 inches.
- The Type-72 is a Chinese-manufactured circular, plastic blast mine. It is designed to damage or destroy a vehicle. For more information, visit http://en.wikipedia.org/wiki/Type_72. Accessed May 15, 2006.
- The Type-72S is a surrogate mine made to be used for the research outlined in this article.
- K.R. Torrance, C.G. Coffey, A.B. Markov and W.S. Myles, “Surrogate AP Mines for Training Deminers and Evaluating Demining Equipment,” Proceedings of the EUDEM2–SCOT2003 (International Conference on Requirements and Technologies for the Detection, Removal and Neutralization of Landmines and UXO), 2003.
- The measurements of water content range from 6 percent to 16 percent for lanes 1 to 5 and 9 percent to 22 p30cent for lane 6.
- So far the reported detection depth is not used in the judgment, but in considering how to use the information is a future work.
- CEN Workshop Agreement, Humanitarian Mine Action—Test and Evaluation—Metal Detectors, CWA 14747, 2003.
- K. Javaraman, A Statistical Manual for Forestry Re search. Food and Agriculture Organization of the United Nations Regional Office for Asia and the Pacific, March 1999. http://www.fao.org/documents/show_cdr.asp?url_file=/DOCREP/003/X6831E/X6831E00.htm. Accessed April 12, 2006.
- Homogeneity of variance is the most important assumption in ANOVA. To keep the assumption, in the case that the observations are proportions or percentages (i.e., probability of detection), derived from frequency data, the observed proportion p may be transformed by angular transformation or logit transformation.
- D. J. Goodenough, K. Rossmann, and L.B. Lusted, “Radiographic Applications of Receiver Operating Characteristics (ROC) Curves,” Radiology, vol. 110, pp. 89–95, 1974.
- C. E. Metz, B. A. Herman, and C. A. Roe, “Statistical Comparison of Two ROC-curve Estimates Obtained from Partially-paired Datasets,” Medical Decision Making, 18, pp. 110–121, 1998.
- K. M. Simonson, “Statistical Considerations in Designing Tests of Mine Detection Systems: I—Measures Related to the Probability of Detection Test Design,” Sandia Report, SAND98-1769/1, Sandia National Laboratories, 1998.
- K. M. Simonson, “Statistical Considerations in Designing Tests of Mine Detection Systems: II—Measures Related to the False Alarm Rate Test Design,” Sandia Report, SAND98-1769/2, Sandia National Laboratories, 1998.
- One square meter is equivalent to 1.2 square yards.
R&D Office for Supporting Antipersonnel Mine Detection and Removal Activities
Japan Science and Technology Agency
10F, 1-32-12, Higashi
Shibuya-ku, Tokyo 150-0011 / Japan
Tel: +81 3 5778 2001
Fax: +81 3 5778 5020
Web site: http://www.jst.go.jp/kisoken/jirai/EN/index-e.html
R&D Office for Supporting Antipersonnel Mine Detection and Removal Activities
Japan Science and Technology Agency
Tokyo Denki University