penetrating radar (GPR) mounted on robotic vehicles. ... Gryphon. 2. TEST AND
EVALUATION RESULTS. In constructing the lanes, all the original soil is ...
Test and Evaluation of Anti-personnel Landmine Detection Based on Vehicle-mounted GPR Systems Jun ISHIKAWA*a, Mitsuru KIYOTAa, and Katsuhisa FURUTAb a
Japan Science and Technology Agency b
Tokyo Denki University 1. INTRODUCTION
This paper shows results of a test and evaluation for anti-personnel landmine detection systems using ground penetrating radar (GPR) mounted on robotic vehicles. Six research teams from universities and industries joined the test and evaluation managed by the Japan Science and Technology Agency (JST). Sensor systems evaluated here have been developed since October 2002 under the program of “Research and Development of Sensing Technology, Access and Control Technology to Support Humanitarian Demining of Anti-personnel Mines” funded by the JST, the competent authority of which is the Ministry of Education, Culture, Sports, Science and Technology (MEXT). Four systems using both GPR and metal detector (MD) were evaluated (Figs. 1-3). Concept of the developed systems is to make no explicit alarm and to provide operators with clear subsurface images. This means that the decision-making whether or not a shadow in the image is a real landmine is entirely left to the operator like medical doctors can find cancer by reading CT images. To evaluate these kinds of systems, a series of tests have been conducted from 8 February 8 to 11 March 2005 in Sakaide City, Japan. Six test lanes were constructed using more than 200 landmine surrogates. Since operators’ preknowledge of the locations of buried targets significantly influences the detection results of our systems, all the 6 lanes are designed to be suitable for blind tests.
Fig. 1. Mine hunter vehicle (MHV) and its sensors (from left to right, MHV#1 and MHV#2).
Fig. 2. Advanced mine Sweeper (AMS).
Fig. 3. Gryphon.
2. TEST AND EVALUATION RESULTS In constructing the lanes, all the original soil is removed with a width of 2m until a depth of 0.5m in vertical section, and the lanes were filled with controlled non-mineralized soil. The width of test lanes is 1m, and mine surrogates were buried up to a depth of 0.3m. Feature of each lane is summarized as follows: -
*
the lanes #1, #2, and #3 lanes are 15m long with flat surface (Fig. 4),
[email protected]; phone +81 3 5778-2001; fax +81 3 5778-5020
-
the lane #4 is 20m long, the surface of which has 15 bumps with a height of 10cm and a diameter of 60cm (Fig. 5), and small stones are mixed into the controlled soil,
-
the lane #5 simulates minefields in post-clearance inspection after mechanical demining, the soil of which is stirred and not packed, and
-
the lane #6 is wet, where 10 liters of water per 1m2 is sprinkled 1 hour before the test starts.
Fig. 4. The lane #3 (under construction).
Fig. 5. The lane #4 (completed).
According to the experimental designs in Tables 1 and 2, more than 200 mine surrogates were buried at the specified depths and angles. As a test result, for example, it has been shown in the lane #2 (Fig. 6) that a metal detector can clearly image a pair of Type72 surrogates buried flush (Fig. 7), and that a ground-penetrating radar can display PMN2 surrogates at a depth of 20cm (Fig. 8). Note that these PMN2s were not able to be detected by using only a metal detector. Tables 1 and 2 show the details of experimental designs. Averages of probability of detection (PD) of 4 testees, each one of which was selected from every system, are plotted in Figs. 9 and 10 compared with the results using only a metal detector. These results showed that probability of detection for targets in deeper levels than 10cm can be improved by combining GPR with MD. On the other hand, as also shown in Figs. 9 and 10, some of GPR+MD results in shallow levels were worse than those of MD. This is because sensor height to the ground controlled by robot arms is higher than that of manual scanning of MD, and this can be improved by fixing a robotic part. Through the test and evaluation, many lessons have been learned, some of which are listed below: -
Provided calibration area should have contained mine surrogates for all the levels of factors. Coaching a typical image for each level would much improve the detection rate.
-
In some case, for example testee #1, high PD have been accompanied by high false alarm rate around 30times/1m2, and it was also proven that confirming the source of GPR false alarm is much more difficult than those of MD, i.e., metal fragments. Therefore, another performance index to penalize these GPR false alarms will be needed.
The most important thing is to use these technologies to improve mine detection efficiency and reduce minefields, and the mine detection systems must be robust, simple and highly cost-effective. From the viewpoint of this, the next step of the project is field tests to evaluate these features in some mine-affected countries. 1 0.5 0 0
1
2
3
4
5
6
7
8
9
10
1
0cm
20cm
0.5 0 10
11
12
13
14
15
Fig. 6. Ground truth of the lane #2: ** shows a pair of Type72 and ∇ shows PMN2.
Fragment
Fig. 7. Detection image from metal detector.
Fig. 8. Detection image from GPR.
2
Table 1. Experimental Design I and the concrete plan to bury mine surrogates: four factors are evaluated: for example, “M14” is Level #1 of Factor A (Target type). Number of Factor 1 2 3 4 Level A: Target type B: Target depth C: Soil condition D: Target angle
4 4 4 2
Experimental run 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Target type
M14 0cm Flat Vertical
Target depth
M14 PMN2 Type72 Type72-S M14 PMN2 Type72 Type72-S M14 PMN2 Type72 Type72-S M14 PMN2 Type72 Type72-S
0cm 10cm 20cm 30cm 10cm 0cm 30cm 20cm 20cm 30cm 0cm 10cm 30cm 20cm 10cm 0cm
PMN-2 10cm Wet Level
Soil condition
TYPE72 20cm Stirred
Target angle
Flat Wet Stirred Rough Stirred Rough Flat Wet Rough Stirred Wet Flat Wet Flat Rough Stirred
Number of target
Vertical Vertical Level Level Level Level Vertical Vertical Vertical Vertical Level Level Level Level Vertical Vertical
7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7
TYP72-S 30cm Rough
Lane used Lane #3 Lane #6 Lane #5 Lane #4 Lane #5 Lane #4 Lane #3 Lane #6 Lane #4 Lane #5 Lane #6 Lane #3 Lane #6 Lane #2 Lane #4 Lane #5
1
Probability of detection
0.9 0.8 0.7
Testee #1 Testee #2 Testee #3 Testee #4 MD
0.6 0.5 0.4 0.3 0.2 0.1 Level
Vertical
Rough
Stirred
Wet
Flat
30cm
20cm
10cm
0cm
TYPE72-S
TYPE72
PMN2
M14
0
Level of factor Fig. 9. Averages of PD for Experiment I in Table 1: testees #1, #2, #3 and #4 respectively from AMS, MHV#2, Gryphon, MHV#1.
3
Table 2. Experimental Design II and the concrete plan to bury mine surrogates: three factors are evaluated: for example, “20cm” is Level #3 of Factor A (Target depth). Factor
Number of Level
1
2
3
4
4 2
0cm Flat
10cm Rough
20cm
30cm
2
15cm
> 50cm
A: Target depth B: Soil condition C: Distance to adjacent target
Target angle: Level, Target: Type72-S
Experimental run
Target depth
1 2 3 4 5 6 7 8
Distance to adjacent target
Soil condition
0cm 10cm 20cm 30cm 0cm 10cm 20cm 30cm
Flat Rough Rough Flat Rough Flat Flat Rough
Number of target
15cm 15cm 15cm 15cm > 50cm > 50cm > 50cm > 50cm
14 14 14 14 14 14 14 14
Lane used Lane #2 Lane #4 Lane #4 Lane #1 Lane #4 Lane #3 Lane #1 Lane #4
1
Probability of detection
0.9 0.8 0.7
Testee #1 Testee #2 Testee #3 Testee #4 MD
0.6 0.5 0.4 0.3 0.2 0.1 >50cm
15cm
Rough
Flat
30cm
20cm
10cm
0cm
0
Level of factor Fig. 10. Averages of PD for Experiment II in Table 2: testees #1, #2, #3 and #4 respectively from AMS, MHV#2, Gryphon, MHV#1.
4
