Shape Reconstruction Using a Mobile Robot for

0 downloads 0 Views 2MB Size Report
is positioned above a target object and dense measurements are taken. However such ... robot faces issues due to vegetation, unevenness due to small holes or hills .... 4th and 5th joints allow setting the horizontal alignment of the metal ... of the 3D map of the environment created by the robot using the LRF and the tilt unit.
Shape Reconstruction Using a Mobile Robot for Demining and UXO Classification Sedat Dogru1 and Lino Marques1 Abstract— Metal detectors are widely used to detect and localize land mines, as well as metallic clutter that causes many false alarms. These false alarms are handled by manual inspection with a prodder or using extra features like size and depth as well as fusion with other sensors like ground penetrating radar or chemical sensors. Directly shape itself has never been used, mainly due to infeasibility of the results of the few existing studies, which require controlled environments and dense sampling of the target objects. In this paper, we propose a new method for shape reconstruction that works with sparse data, and show its feasibility using field data collected by a mobile robot equipped with a commercial pulse induction metal detector carried by a custom 2DoF arm. This paper also describes the method employed to sweep natural terrains with the mobile manipulator and provides results of imaging a set of different metallic objects.

I. INTRODUCTION Landmines and unexploded ordnance (UXO) are some of the major humanitarian problems threatening citizens across dozens of developing countries. Using robots in landmine and UXO detection seems a natural fit, taking into account the dangerous, tedious, and time consuming nature of the job. Therefore researchers have either customized existing mobile robots for this purpose [1], or have built custom demining robots from scratch [2]–[6]. Those robots are equipped with various types of landmine/UXO detectors, like metal detectors (MD), ground penetrating radars (GPR), infrared detectors, or detectors for vapors from explosives. However, metal detectors are the most common because of their sensitivity to metals, low cost and quite mature and reliable technology compared to other type of sensors. Metal detectors are in broad use for this task since WWII. Despite these advantages, metal detectors have a significant disadvantage: they give a warning for every metal they detect, causing many false alarms, further increasing the cost of already expensive demining. In order to mitigate those false alarms some researchers fuse metal detector data with data from different type of sensors like GPR [7] or explosive vapor detector [8]. Others utilize various methods like kNN [9], [10], or feature descriptors [11] to discriminate between different types of detections of the same sensor. Shape detection from metal detector data has also been studied [12]– [14], though not in a demining context. The prior work on This work was partially carried out in the framework of TIRAMISU (www.fp7-tiramisu.eu). This project was funded by the European Community’s Seventh Framework Program (FP7/SEC/284747). 1 Sedat Dogru and Lino Marques are with Institute of Systems and Robotics, Department of Electrical and Computer Engineering, University of Coimbra, 3030-290 Coimbra, Portugal {sedat,

lino}@isr.uc.pt

this area focus on shape reconstruction in lab environments under controlled conditions, where a gantry scanner system is positioned above a target object and dense measurements are taken. However such an approach is not feasible in outdoor demining missions. In this paper we show that a field robot equipped with a commercial metal detector can be used to reconstruct shape using sparse measurements. We also propose a new method for the solution of the problem of shape reconstruction using a metal detector, relaxing the strict metal detector transfer function requirements of the previous studies. Although our method works properly with sparse measurements it still benefits from a platform that is properly controlled because of the following aspects: Response of a metal detector depends on the distance as well as the material, shape and size of the target object. And sensitivity of the metal detector decreases with increasing distance to the target. In order to optimize detection, the metal detector has to be as close as possible to the scanned surface, which may not be an easy task in rough terrain. On rough terrain the robot faces issues due to vegetation, unevenness due to small holes or hills, stones on the surface, all requiring adapting the manipulator to the surface. Those may also cause issues since they may change the attitude of the robot, sometimes suddenly, and as a result leaving the metal detector far off the surface or in contact with it. In order to compensate for this, in this paper we continuously construct the profile of the terrain using a tilted laser range finder (LRF), which is widely used on robots for obstacle avoidance. The tilt unit constructs the 3D representation of the surrounding environment, allowing the robot evaluate the transversable areas, detecting not only large obstacles but also smaller obstacles that could hinder its motion, like bushes, small rocks etc. The 3D scan of the environment is used in this work to construct the height profile of the points that will be swept in the next step, which is afterwards followed using a two stage controller for the 2DoF robotic arm holding the metal detector. The robotic arm is controlled so that the end effector is sweeping at a constant speed, and also it is passing closely over the target surface. This paper has three main contributions. Firstly, a novel method to the shape reconstruction problem using a metal detector is proposed. Secondly, shape reconstruction using a mobile robot is shown. Finally, a complementary problem, sweeping a rough terrain with a metal detector using feedback about the surface from a laser scanner is solved.

II. RELATED WORK Since EMI sensors create eddy currents that in turn create a magnetic field, the targets are identified by a magnetic polarizability matrix, which correlates the applied voltage to the measured one. Dekdouk et al. [15] propose a deterministic nonlinear optimization method to estimate the magnetic polarizability tensor using a coil that passes over the target, and show in simulation that the tensor can be estimated with 12% of error for a sphere and an elsie ring. Bell et al. [16] scan a target area in three orthogonal directions and use the correlation between the eigenvalues of the magnetic polarizability matrix to identify whether an object is a long slender object (UXO like), or a fat one. Norton et al. [17] use the same principle and show that for a wide range of frequencies the ratio of the eigenvalues is smaller than one for long symmetric objects and it is greater than one for flat or irregular shaped objects. Fails et al. [9] work in the frequency domain and instead of relying on the magnetic polarizability tensor use a four parameter frequency representation which is used with a kNN to distinguish UXOs from the rest. Tantum et al. [10] similarly use a kNN to distinguish UXOs, but they use Discrete Spectrum of Relaxation Frequencies (DSRF), which represent the response of the object as the sum of discrete relaxation frequencies, as features. Das et al. [18] study the response of a metallic sphere and a metallic prolate spheroid to a large metal detector at various depths and orientations reporting partial success. So there are various studies that try to classify UXOs and landmines, as well as studies that try to estimate the depth and orientation of the targets. However, there are just few papers that discuss shape reconstruction for UXOs. The first paper is by Bruschini [12], who utilizes a commercially available imaging metal detector (Ferroscan RV 10) to detect and visualize UXOs/landmines. The imaging metal detector is used normally to locate and display metal bars inside concrete, and operates by measuring the gradient of the induced magnetic field using an array of sensors. Druyts et al. [13] study shape reconstruction in detail using a pulse induction metal detector, deriving target models for different types of targets and then reconstructing those targets from experimental data (ball, and wires of various shape: straight, X-like, flag-like and rectangle). Then they use the matching target function to reconstruct the target at various depths, varying from 3 cm to 8 cm. Their experimental setup consists of an X-Y table sampling every 5 mm. Kruger and Ewald [14] study the same problem using a two frequency continuous wave metal detector and reconstruct the metallic parts of a low metal anti-tank mine, which has three balls placed at some distance from each other, as well as an aluminum ball and a steel ball placed at depths of 2 cm, 4 cm and 8 cm. Their setup also consists of an X-Y table collecting data densely with precise position information. The transfer function of a metal detector, which is defined as the magnetic field intensity it generates at a given height across the surface, changes both in height and magnitude. It also changes depending on the target shape, therefore

Tilt Unit The Arm Metal Detector

LRF Linear Actuator Pitch Joint Sweep Joint

Fig. 1. Husky, the robot used in this work, with its components referred in the text labeled.

both Druyts et al. [13] and Kruger and Ewald [14] assume that the scanned object is known, and hence use a proper transfer function to reconstruct the shape, which is not feasible in general. Our approach in this paper is based on approximating the transfer function of the metal detector, and hence obtaining a coarse estimate of the shape of the target, providing valuable clues to a demining or UXO search mission. Various research groups have utilized robotic arms in their robot platform for scanning for land mines. Santos et al. [5] as well as Montes et al. [19] have used robotic arms to carry a metal detector on their hexapod robot, and use infrared sensors that are mounted next to the metal detector to track the surface as well as obstacles at the height of the metal detector. Kaneko et al. [20] use stereo vision to map the terrain for trajectory optimization of their robotic arm mounted on a mobile platform. However, vision is susceptible to changes in light in the environment. Infrared sensors next to the metal detector may interfere with or be interfered by the metal detector due to electromagnetic noise each creates. Another disadvantage of having sensors so close to the end effector is, the controller doesn’t have the chance to get the global picture, and hence plan the trajectories upfront accordingly. III. MATERIALS AND METHODS Since description of the methods strongly depends on the hardware used, this section starts with a brief description of the hardware employed. This is followed by the method which consists of 4 main parts: First, there is robot motion control, which is responsible for adjusting the speed of the robot as it searches for a target or as it sweeps over a target to estimate its shape. The second part is terrain profile reconstruction, whose output will be used by the third stage, namely control of the robotic arm, both for shape reconstruction and target search. And the final part is estimation of the shape using the collected data. A. The Robot In this work, a heavily customized Husky skid steered mobile robot from Clearpath was used (Fig. 1). The original base was extended with a bridge hosting some sensors, like a pair of stereo cameras, a laser with a tilt unit, dual

Trajectory of the metal detector 0.6

5

0.4

0.4

3

0.2

0.2

0 −0.2

2 1

0

0 −4

−0.2

−2

−0.4 1

1.5 x (meters)

(a)

2

0 x(m)

2

4

2

4

(b) 1

1.5 x (meters)

2

(b)

1 z(m)

−0.4

4 y(m)

y (meters)

y (meters)

Trajectory of the metal detector 0.6

Fig. 2. (a) The ideal trajectory of the metal detector as the arm sweeps between −30◦ and +30◦ and the robot moves to the right at constant speed, without stopping (b) The trajectory of the metal detector as the arm sweeps between −30◦ and +30◦ however this time the robot stops before each scan.

−2

0 x(m)

(c)

z(m)

GPS antennas backed by a base station providing position accurate at centimeter level as well as heading. The robot was equipped with a homemade 5 revolute joint arm. The first joint is controlled by a brushless DC motor which is connected to it through a belt that allows sliding, preventing any harm to the driving motor or the arm itself if it hits an obstacle. This joint is monitored by a high resolution absolute encoder providing 1024 points per revolution. The second and third joints are connected to each other in a parallelogram, to keep the end effector parallel to the robot’s base, and this parallelogram is controlled by a linear motor which controls the height of the end effector. The linear actuator is supported by an air shock to add compliance in this axis, further improving shock resistance of the arm in case it hits an obstacle. This joint is also monitored by an absolute encoder providing 16384 points per revolution. The 4th and 5th joints allow setting the horizontal alignment of the metal detector with respect to the arm (yaw) and roll angle of the end effector respectively, though they are set manually, without any electronic control. This configuration effectively lets 2 DoF motion of the arm keeping the end effector parallel to the robot’s base.

0 −1 −4

1.2 1 0.8 0.6 0.4 y(m)

−0.6

−0.8 −1.5

−1

−0.5

0 0.5 x(m)

1

(d) (a) Fig. 3. (a) A trail, a typical place to lay mines (b), (c) Top and back view of the 3D map of the environment created by the robot using the LRF and the tilt unit. Each curve corresponds to a single laser scan, and its color indicates its height, blue being the deepest. The magenta and red dots on the center of (b) indicate spots that will be swept in the next sweep. (d) The trajectory to be followed by the end effector to avoid hitting the ground but staying as close as possible.

representation of the environment (Fig. 2b), instead of one that has more data points at the edges, and less in between (Fig. 2a). Upon finishing sweep for shape, the robot resumes sweep for mine mode. Depending on the report about the shape, and some data from other sensors or other features, the robot can decide if this detection was a mine and avoid it, or just ignore and pass over it. Thanks to these two different speed modes the robot can cover a given area more quickly, increasing efficiency of the overall process. C. Terrain profiling

B. Robot Motion Control The purpose of this part is to drive the robot along paths dictated by a coverage path planning algorithm at two different speeds. While the metal detector isn’t detecting any anomaly, the robot will be moving and sweeping at the maximum speed allowed by the metal detector, which has a non-zero response time. The maximum speed will also be restricted by the allowable gap size in the sweep pattern that emerges (Fig. 2a). This gap size will depend on the size of the objects that are searched for. As soon as a meaningful anomaly is detected, the robot raises an alarm, and switches to slow speed mode, moving ahead at small discrete steps, and sweeping slowly in between these steps. As has been shown in the literature, shape reconstruction requires a dense set of data points, and in order to collect this dense set the robot has to scan the area of interest by moving slowly as well as moving its sensor head slowly. Movement in discrete steps, instead of continuously at a low speed, allows the metal detector construct a uniform

In order to properly follow the terrain and properly acquire data with the metal detector, the terrain profile should be known, either in advance or in real-time, while the robot covers the minefield. In this paper, we take a real-time approach, which allows taking into account changes in the terrain that may have happened since the last planning phase. The 3D profile of the terrain is constructed using a tilt mounted LRF (Figs. 1, 3a). Scans of the LRF are fused with the tilt angle of the tilt unit, providing a 3D point cloud of the environment (Fig. 3). In order to simplify the control of the arm and to prevent introduction of new errors due to inaccuracies in the global localization system, all the LRF measurements are transformed from the LRF tilt unit’s frame of reference, referred as L, to the frame of reference of the base of the arm, referred as A in this text. This way a more accurate terrain representation can be constructed instantaneously with respect to the base of the arm, and hence appropriate control action can easily be estimated and applied.

The transformation sequence of the point clouds starts with the optical center of the LRF, which is translated with respect to the tilt joint (x1 , y1 , z1 ). The tilt joint can rotate around the pitch axis (φL ), and it is translated with respect to the 1st joint of the arm (x2 , y2 , z2 ). The 1st joint of the arm can rotate around the yaw axis (θA ), and is connected to the 2nd joint of the arm by a translation (0, 0, z3 ). The second joint of the arm is capable of rotating around the pitch axis (φA ). The metal detector is fixed to the arm, further away at the fifth joint, but since all those joints are fixed, the metal detector can be taken to be fixed at (0, 0, z4 ) with respect to the arm’s pitch controller. So, a point measured by the LRF and expressed as w = (rL cos θL , rL sin θL , 0, 1)0 in homogeneous coordinates in its frame of reference, where r is the distance to the point and θL is the angular displacement with respect to the LRF’s x-axis, is transformed to the arm’s 2nd joint’s reference frame by ArmBase

TTilt

Tilt

TLaser

Metal Detector

Terrain Profiler

Robot Motion Controller

Arm Controller

Lift Controller

Sweep Controller

Fig. 4. A global view of the control architecture used on the robot. The robot motion controller, responsible from driving the robot on different paths depending on a coverage planner, and at different speeds depending on the output of the metal detector. Arm controller decides on the sweep speed depending on the speed of the motion controller, dictating this to the sweep controller, and also dictates the lift controller the height it should keep in the next instant.

Laser

TWorld w = ArmPitchJoint TArmPitchJoint TWorld w0

ArmBase

(1)

where T is the transformation matrix from the reference frame defined by the lower index to the upper index. After solving the above equation, distance, yaw and pitch of the target point with respect to the arm base is obtained as (rA , θA , φA ), which directly correspond to the control variables of the arm. Solution of equation 1 gives x = rL cos θL + x1 + x2 y = rL cos φL sin θL + y1 cos φL − z1 sin φL + y2 z = rL sin φL sin θL + y1 sin φL + z1 cos φL + z2 + z3 (2) which are coordinates of the point cloud transferred to the arm in Cartesian coordinates, giving us the following for (rA , θA , φA ) p rA = x2 + y 2 + z 2     (3) −x z θA = arctan , φA = arcsin y rA After this transformation these points are filtered, so that only the points that will be swept in the current iteration are left out using the following constraint, which specifies the boundary conditions on the region that will be swept in the next step RArm − δr < rA cos φA < RArm + δr min max θArm < θA < θArm

(4)

min max where θArm , θArm , RArm and δr are the minimum and maximum sweep angles defined for the next iteration, length of the arm and half length of the metal detector, which is connected to the arm at its middle, respectively. Afterwards, the left points are further processed by choosing the maximum height for each Aθi for the points that are under the end effector using

g( Bθi ) = max zi ( Bθi , Bφj ) j

Mission Planner

(5)

which chooses the maximum height along the line at Bθi . Finally the pitch angle of the arm at a given polar angle as it sweeps over a rough surface can be calculated using: g( Bθ ) − z4 (6) RArm where z4 was the translation of the end effector from the arm. Note that this is the minimum height to be kept by the metal detector at that point for it to be barely touching that point, ignoring its own width. In order to account for the width of the metal detector, the obtained curve should be convolved with a pulse kernel that has width of the metal detector at least, this way it is made sure that no matter at which point the end effector is, it won’t be hitting anything. Φ = Φ( Bθ ) =

D. Height Control of the Arm Although the arm on the robot has 5 joints, only the first two joints are controllable, allowing control in the yaw axis (θ) for sweep, and then in the pitch axis (φ) for lift respectively. Arm, carrying the metal detector at its end point, requires proper control both in the pitch and the yaw axes. Control of the pitch axis is necessary to ensure proper surface following at close range and therefore requires control of the position, whereas the yaw axis requires control of its speed both to ensure sweeping of the surface at a uniform rate and to give the pitch controller enough time to achieve its control goal, i.e. rise or sink to follow surface irregularities. Fig. 4 shows an overview of the control block diagram for the arm. The arm controller is expected to generate proper sweep speed and lift position controller commands for the corresponding blocks. It does this by using the outputs of section III-C, which provides the appropriate lift angle as a function of speed angle for the next sweep, and also using outputs of the robot motion controller, which determines the speed the arm should be sweeping. Note that this speed will depend on the metal detector output, which will be high if the metal detector is not detecting anything, and low as soon as it detects something. Another thing that should be taken into

x 10

Metal detector response

The current passing through the coils of a pulse induction metal detector creates a magnetic field, which causes eddy currents on close metal objects as soon as it is turned off. These eddy currents create a magnetic field which is detected by the coils. All those interactions can be calculated using Maxwell’s Equations. However due to the geometry of the coils and shapes of the targets analytical solutions may not be possible. 1) Metal Detector Response: The employed Vallon VMP3 pulse induction metal detector does not provide directly the voltages induced in its coils, instead it processes the detected voltages and provides the output as three channels. The first two channels are more sensitive compared to the third channel, however as was found out during the experimental runs, and as was confirmed by Vallon, the values reported from the first two channels are not generally consistent with the distance to the target. For some targets it was observed that the reported values may change unpredictably depending on the target object, like decreasing as the coil gets closer to the target, leaving to us the less sensitive third channel to work with any type of metallic object in a predictable manner. This behavior can be seen in Fig. 5 where the metal detector response as it sweeps over the object by passing directly over the center (Fig. 5a) and 26 cm away from the center (Fig. 5b) is presented. 2) Model of the Imaging Process: In this paper, we model the measurement process by a convolution of the coil transfer function and the target, including a noise term, which is expected to include the nonlinearities as well: (7)

where h(x, y) is the coil transfer function, f (x, y) is the target, g(x, y) is the induced voltage, and η is the noise term. Normally the interactions between the coil and the targets are a little more complex involving nonlinearities, however it was

4

1st channel 2nd channel 3rd channel

−0.4

−0.2

0

0.2

0.4

0.6

Magnitude

15 10 5 0

15 10 5 0

x 10

Metal detector response 1st channel 2nd channel 3rd channel

−0.4

(a)

−0.2

0

0.2

0.4

0.6

(b)

Fig. 5. Metal detector response to a target placed at θ = 0.1 rad, as it sweeps over it from different positions, at 10 cm above the object (a) almost over the center of the object, (b) 26 cm away from the center of the object. Note that the data has a different offset for each channel, and this has been filtered out to be able to compare the magnitude of the changes on the same graph.

shown that this convolution process is a good approximation for reconstructing the shape of the target [13], [14]. The original shape of the object can be recovered from equation 7 using a Wiener deconvolution filter, which works in frequency domain and minimizes the mean square error between the estimated and the real function. The Wiener deconvolution filter is given by Fˆ =

E. Image Reconstruction

g(x, y) = h(x, y) ∗ f (x, y) + η

4

Magnitude

account is the height variations of the region to be swept. If the height variation is too large, than the controller has to decrease the speed of the arm even further, to allow time for the lift controller react on the way. The sweep controller aims to achieve a trapezoidal speed profile for the sweep speed of the arm, reducing shaking and providing a uniform sampling of the surface under it, both in space and signal strength. Although the underlying hardware allows changing the accelerations of the motor, and hence achieve the trapezoidal speed profile faster, it is not preferred specially during measurement for the shape phase, since this will cause long lasting vibrations in the arm, degrading quality of the data. Lift controller has to be more reactive compared to the sweep controller. It’s required to change height of the end effector, by changing the pitch angle of the arm, continuously. This is achieved using a PI controller. The arm has speed constraints due to the response time of the detector, its sensitivity and its sampling frequency. The metal detector in use provides 30 samples per second.

G H∗ |H|2 + Snoise SG

(8)

where Fˆ is the estimated target, and SG and Snoise are the spectral densities of the target and noise signals respectively. However for this reconstruction to work H should be known. 3) Coil Model: Accurate calculation of H is difficult both due to the shape of the metal detector, which is more like an ellipse with vertical edges in our case, and difficulty of calculating accurate analytic expressions even for very simple shapes like a circular coil, which after application of Biot-Savart [21] law lead to elliptic integrals. Therefore in this paper we make two simplifications: First, we assume that the metal detector is circular, with a size equal to the size of the largest undeformed coil fitting in it, calling its radius R. This simplification is also motivated by the rule of thumb by TI1 saying the maximum sensing distance of a coil is dominated by the shortest dimension of the coil. The second assumption constraints the area of interest: We assume that the target is placed at a distance r to the center of the coil and r > R, which allows us to roughly approximate the magnetic field equation of a coil in spherical coordinates centered at the center of the coil by (the derivations were left out for brevity of the text) ~r = sin(θ) H r3

~θ = cos(θ) H r3

(9)

Although these two approximations seem too relaxed, their validity is supported by the experimental results obtained. In this work the target objects are assumed to be a collection of dipoles, and the metal detector coil excites each dipole, whose magnetic moment is M, inducing a voltage of E = HM This voltage will in turn induce a voltage back on the coil, generating a voltage of E = HMH which as a result 1 https://e2e.ti.com/support/sensor/inductive-sensing/f/938/t/295036#Q2

gives

Measured, L 0deg

1 ~r |2 + |H ~ θ |2 ) = 1 = h(x, y, z) = (|H 6 2 r (x + y 2 + z 2 )3 (10) for the voltage induced on the coil. Note that dipoles will also interact with each other, giving rise to further eddy currents, though those will be very small, and hence ignored. All the coefficients are taken to be 1, since they are the same irrespective of the target and its position. The coefficients are fixed for a given coil. And moreover, the metal detector provides us an integrated and scaled version of the measured voltage, instead of directly the voltage. IV. R ESULTS In this work we assume that the soil is cooperative, i.e. it is not magnetizable. In order to simplify the work of estimating the performance of shape reconstruction at different heights, and taking into account our cooperative soil assumption, targets of various shapes were placed on a flat asphalt road in various orientations with respect to the scanning direction. The robot was driven along a line by stopping every 2 cm and sweeping an 80 degrees arc.

Estimated, bb 20 by 25

10

10

20

20

30

30

40

40 10

(a)

20

30

40

10

(b)

Measured, L 45deg

Estimated, bb 28 by 20

Measured, L 90deg

10

10

10

20

20

20

20

30

30

30

30

40

40

40

20

30

40

10

(d)

20

30

40

(e)

20

30

40

10

20

20

30

10

10

20

20

30

30

20

30

40

10

(b)

Measured, I32 45deg

Estimated, bb 20 by 18

Measured, I32 90deg

10

10

10

20

20

20

30

30

30

30

40 20

(d)

30

40

40 10

20

(e)

30

40

20

(f)

30

40

30

40

10

Measured, E 0deg

10

20

30

20

30

40

(c)

Fig. 8. A galvanized steel C strip formed by using 6a and 7a (a) The real shape recorded on camera. (b) The raw data, which is formed by sweeping over the shape at 0 degrees with the metal detector. (c) Estimated shape.

40 10

40

40

Estimated, bb 24 by 9

20

10

30

(c)

10

40

20

20

(b)

40 10

30

40 10

(a) (a)

20

(g)

Estimated, bb 24 by 26

40

30

40

10

(f)

Fig. 7. A galvanized steel L measuring 33cm x 22cm x 0.65cm and strip width of 3cm (a) The real shape recorded on camera. (b) The raw data, which is formed by sweeping over the strip at 90 degrees with respect to the metal detector. (c) Estimated shape. (d) The raw data, which is formed by sweeping over the L at an angle of 45 degrees with the metal detector. (e) The estimated shape (f) The raw data, which is formed by sweeping over the L at an angle of 0 degrees with the metal detector. (g) The estimated shape.

Estimated, bb 8 by 25

10

40

40 10

Measured, C 0deg

Measured, I32 0deg

30

Estimated, bb 25 by 21

10

10

20

(c)

40

(g)

Fig. 6. A galvanized steel strip measuring 32cm x 3cm x 0.65cm. (a) The real shape recorded on camera. (b) The raw data, which is formed by sweeping over the strip at 90 degrees with the metal detector. (c) The estimated shape. (d) The raw data, which is formed by sweeping over the strip at an angle of 45 degrees with the metal detector. (e) The estimated shape (f) The raw data, which is formed by sweeping over the strip at an angle of 0 degrees with the metal detector. (g) The estimated shape.

In this work, shapes made with two different metals were studied. The first set consisted of various shapes formed by a single or many piece galvanized steel stripes of 3 cm width and 0.65 mm thickness. The shapes were a 32 cm long I (Fig. 6), an L (Fig. 7) whose long edge is 33 cm long, and short edge is 22 cm long, as well as a C (Fig. 8a), formed by extending the L shape with the 32 cm long I (Fig. 6), and an E (Fig. 9a) formed by extending the C with a 22 cm long I placed in the middle. The I and L shapes were scanned in three different orientations, namely 90 degrees, 45 degrees and 0degrees with respect to the scanning direction, whereas the C and E were placed only at 90 degrees and 0 degrees. The simplest shape was the I, and as can be seen in Fig. 6c, 6e, and 6g the shape is properly reconstructed from the raw

Estimated, bb 23 by 26

10

10

20

20

30

30

40

40 10

(a)

20

30

40

10

(b)

20

30

40

(c)

Fig. 9. A galvanized steel E strip formed by using 6a and 7a and a 22cm long strip (a) The real shape recorded on camera. (b) The raw data, which is formed by sweeping over the shape at 90 degrees with the metal detector. (c) Estimated shape.

Measured, C 90deg

Estimated, bb 26 by 24

Measured, E 90deg

Estimated, bb 26 by 23

10

10

10

10

20

20

20

20

30

30

30

40

40 10

20

(a)

30

40

30

40 10

20

(b)

30

40

40 10

20

(c)

30

40

10

20

30

40

(d)

Fig. 10. (a) The raw data of 8a, which is formed by sweeping over the shape at 0 degrees with the metal detector. (b) Estimated shape. (c) The raw data of 9a, which is formed by sweeping over the shape at 0 degrees with the metal detector. (d) Estimated shape.

Measured, Loop

Estimated, bb 25 by 26

10

10

20

20

30

30

40

(a)

40 10

20

30

40

(b)

10

20

30

40

(c)

Fig. 11. A stranded copper wire of 5 conductor wires, measuring 5mm in diameter and 132cm in length, forming a rough ellipse. (a) The real shape recorded on camera. (b) The raw data, which is formed by sweeping over the wire loop with the metal detector. (c) Estimated shape.

data given in 6b, 6d, and 6f for 0, 45 and 90 degrees respectively. However, despite the proper estimation of the shape, the reconstructed shape is a significantly diluted version of the original shape. The algorithm is much more successful in reconstructing the L in three different orientations (Fig. 7c, 7e, 7g) as well as the C in two different orientations (Fig. 8c, 10b). Unfortunately the algorithm was not able to reconstruct the inner details of the E, though its outer lines are realistically depicted (Fig. IV). The second metal tested was copper. For this an ellipse was formed by bending a stranded copper wire of 5 conductor wires, measuring 5mm in diameter and 132cm in length (Fig. 11a). Despite the high noise present in the raw data for this specific target (Fig. 11b) compared to others, the proposed method was able to reconstruct a satisfactory shape as seen in Fig. 11c. V. CONCLUSIONS In this paper the problem of shape reconstruction of buried metallic objects using an autonomous mobile manipulator was studied from different aspects. The first aspect is creating a 3D profile of the surface to be scanned with a metal detector, which is achieved by the aid of a LRF mounted on a tilt unit, creating a 3D point cloud. This 3D point cloud was later transferred to the reference frame of the arm, and the cloud was filtered so that only reachable points are left, and using these reachable points a height map was created. This map was later followed by the metal detector through control of the arm carrying it. Control of the arm consisted of two main components, constant speed drive along the sweep axis and position control along the height axis. Validation of uneven surface scanning is shown in the video attachment. Finally a new method was presented to construct shape from metal detector readings. It was shown that coarse shapes of different objects placed in different orientations can be successfully estimated, even when the input data is noisy. R EFERENCES [1] G. Cabrita, R. Madhavan, and L. Marques, “A framework for remote field robotics competitions,” in Autonomous Robot Systems and Competitions (ICARSC), 2015 IEEE International Conference on, April 2015, pp. 192–197. [2] J. Nicoud and M. Habib, “The pemex-b autonomous demining robot: perception and navigation strategies,” in Intelligent Robots and Systems 95. ’Human Robot Interaction and Cooperative Robots’, Proceedings. 1995 IEEE/RSJ International Conference on, vol. 1, Aug 1995, pp. 419–424 vol.1.

[3] M. Y. Rachkov, L. Marques, and A. T. de Almeida, “Multisensor demining robot,” Autonomous Robots, vol. 18, no. 3, pp. 275–291, 2005. [4] P. Santana, C. Cˆandido, P. Santos, L. Almeida, L. Correia, and J. Barata, The Ares Robot: Case Study of an Affordable Service Robot. Berlin, Heidelberg: Springer Berlin Heidelberg, 2008, pp. 33–42. [5] P. G. de Santos, J. Cobano, E. Garcia, J. Estremera, and M. Armada, “A six-legged robot-based system for humanitarian demining missions,” Mechatronics, vol. 17, no. 8, pp. 417–430, 2007. [6] E. F. Fukushima, M. Freese, T. Matsuzawa, T. Aibara, and S. Hirose, “Humanitarian demining robot Gryphon-current status and an objective evaluation,” International journal on smart sensing and intelligent systems, vol. 1, no. 3, pp. 735–753, 2008. [7] C. Bruschini, B. Gros, F. Guerne, P.-Y. Pice, and O. Carmona, “Ground penetrating radar and imaging metal detector for antipersonnel mine detection,” Journal of Applied Geophysics, vol. 40, no. 13, pp. 59–71, 1998. [8] J. Prado, G. Cabrita, and L. Marques, “Bayesian sensor fusion for landmine detection using a dual-sensor hand-held device,” in Industrial Electronics Society, IECON 2013 - 39th Annual Conference of the IEEE, Nov 2013, pp. 3887–3892. [9] E. B. Fails, P. A. Torrione, W. R. Scott, Jr., and L. M. Collins, “Performance of a four parameter model for modeling landmine signatures in frequency domain wideband electromagnetic induction detection systems,” in Proc. SPIE 6553, Detection and Remediation Technologies for Mines and Minelike Targets XII, vol. 6553, 2007, pp. 65 530D–65 530D–8. [10] S. L. Tantum, K. A. Colwell, W. R. Scott, P. A. Torrione, L. M. Collins, and K. D. Morton, “Sparse model inversion and processing of spatial frequency-domain electromagnetic induction sensor array data for improved landmine discrimination,” in Proc. SPIE 8709, Detection and Sensing of Mines, Explosive Objects, and Obscured Targets XVIII, vol. 8709, 2013, pp. 87 091E–87 091E–8. [11] S. Larionova, L. Marques, and A. T. de Almeida, “Multi-stage sensor fusion for landmine detection,” in 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, Oct 2006, pp. 2943–2948. [12] C. Bruschini, “Evaluation of a commercial visualizing metal detector for UXO/mine detection: the hilti ferroscan system,” in International Workshop on Sustainable Humanitarian Demining (SusDem 97), 1998, pp. 6–18. [13] P. Druyts, L. Merlat, and M. P. J. Acheroy, “Modeling considerations for imaging with a standard metal detector,” Proc. SPIE, vol. 4038, pp. 1431–1451, 2000. [14] H. Kruger and H. Ewald, “2D image reconstruction from blurred and disturbed multi parameter measurements at the example of an inductive metal detector,” in Sensors, 2010 IEEE, Nov 2010, pp. 1325–1328. [15] B. Dekdouk, L. Marsh, D. Armitage, and A. Peyton, “Estimating magnetic polarizability tensor of buried metallic targets for land mine clearance,” in Ultra-Wideband, Short-Pulse Electromagnetics 10, F. Sabath and E. L. Mokole, Eds. Springer New York, 2014, pp. 425– 432. [16] T. Bell, B. Barrow, and N. Khadr, “Shape-based classification and discrimination of subsurface objects using electromagnetic induction,” in Geoscience and Remote Sensing Symposium Proceedings, 1998. IGARSS ’98. 1998 IEEE International, Jul 1998, pp. 509–513 vol.1. [17] S. Norton, I. Won, and E. Cespedes, “Ordnance/clutter discrimination based on target eigenvalue analysis,” Subsurface Sensing Technologies and Applications, vol. 2, no. 3, pp. 285–298, 2001. [18] Y. Das, J. E. Mcfee, J. Toews, and G. Stuart, “Analysis of an electromagnetic induction detector for real-time location of buried objects,” Geoscience and Remote Sensing, IEEE Transactions on, vol. 28, no. 3, pp. 278–288, May 1990. [19] H. Montes, L. Mena, R. Fern´andez, J. Sarria, and M. Armada, “Inspection platform for applications in humanitarian demining,” in 18th International Conference on Climbing and Walking Robots and the Support Technologies for Mobile Machines. Assistive Robotics, Sept 2015, pp. 446–453. [20] A. M. Kaneko, M. Marino, and E. F. Fukushima, “Humanitarian demining robot Gryphon: New vision techniques and optimization methods,” in Intelligent Robots and Systems (IROS), 2010 IEEE/RSJ International Conference on, Oct 2010, pp. 228–233. [21] D. Cheng, Field and wave electromagnetics. Addison-Wesley Publishing Company, 1989.