Capacitive touch sensing : Signal and image processing algorithms

Zachi Baharava and Ramakrishna Kakaralab

a Corning b School

West Technology Center, Corning Incorporated, 1891 Page Mill Rd., Palo Alto, CA of Computer Engineering, Nayang Technological University, Singapore
ABSTRACT

Capacitive touch sensors have been in use for many years, and recently gained center stage with the ubiquitous use in smart-phones. In this work we will analyze the most common method of projected capacitive sensing, that of absolute capacitive sensing, together with the most common sensing pattern, that of diamond-shaped sensors. After a brief introduction to the problem, and the reasons behind its popularity, we will formulate the problem as a reconstruction from projections. We derive analytic solutions for two simple cases: circular nger on a wire grid, and square nger on a square grid. The solutions give insight into the ambiguities of nding nger location from sensor readings. The main contribution of our paper is the discussion of interpolation algorithms including simple linear interpolation , curve tting (parabolic and Gaussian), ltering, general look-up-table, and combinations thereof. We conclude with observations on the limits of the present algorithmic methods, and point to possible future research. Keywords: Capacitive Touch Sensor, Multi Touch.

1. INTRODUCTION
Touch is one of the 6-senses humans use to feel and interact with their environment, and has long been a key component in human-computer interactions. It has come to the forefront in recent years with the explosion in electronic devices using it, most notably mobile phones. Prior to the development of touchscreen phones, touch sensing was used in many places including computer track-pads. With such a wide range of applications, there are many dierent requirements and constraints on the sensing system.1 For example, some systems need to sense both human-nger, stylus, pencil, or pen, while others are tailored only for sensing sub-sets of those. Some systems, like smart-phones, have to be clear, meaning one should be able to see the content (usually an LCD screen) underneath the systems, while other systems can be opaque (like trackpads). Yet some systems need to be able to determine the locations of multiple ngers touching the screen, while other systems need merely to determine whether a button was pressed or not. And of course, there are dierences in the size-constraints, power consumption, accuracy requirements, cost, and so on. Moreover, the wide range of requirements gives rise to dierent methods to sense touch. The most common technique today is still resistive touch sensing, which relies on two conductive layers separated by a small space, and when the user touches, a conductive path is created between the two layers. Note that this method relies on the pressure created when the user touches, and not on the presence of the nger/object per-se. Other methods to sense touch include optical methods, acoustic-waves, resistive-nodes, and various types of capacitive sensing: mutual, absolute, and surface. For a general review of touch technologies, which is updated frequently, a very good resource is Je Walkers presentation.2 For a treatment of the history of capacitive sensors see Baxters classic book.3 In this paper we are focusing on absolute capacitance touch sensing, used to determine the location of one nger over a sensing-area. This method used to be the most common method for capacitive sensing, though in the last three years the focus has moved in many segments to mutual capacitive sensing, which enables a more detailed sensing of the space by creating a capacitance-image of the area sensed. Mutual capacitance is
Further author information: Zachi Baharav - [email protected] , Ramakrishna Kakarala [email protected]

the method, for example, used in the iPhone c and in other devices capable of multi-touch sensing. Never the less, absolute capacitance is still heavily used in many applications where a less-expensive solution is required, or where there is no real need for capacitive imaging capability. As the nal goal of the sensing in our case is to determine a nger-touch coordinate, we focus in this paper on various methods to calculate the coordinate given the sensor data. The structure of the paper is thus as follows: in Sec. (2) we describe the general system we will be using, and lay down the terms to be used in the sequel. We touch briey on projection-approach in Sec. (3) , and in Sec. (4) we analyze a special case, where the sensor is composed of thin-wires. This will lend itself to analytic solution. We also derive an analytic result for square ngers in Sec. 4.2. Sec. (5) discusses dierent methods for calculating the coordinates of the nger, and Sec. (6) describes the simulations and the results. We close with Sec. (7) , and point to possible future work on the subject.

2. PROBLEM FORMULATION
In this section we describe an absolute capacitance sensing system, which is used to sense one nger, and determine its coordinate (location) over a rectangular area. Absolute capacitance sensing systems can be divided in general into three main ingredients: physical sensor, electronic measurement, and signal processing. Touch Sensor

Physical Sensor

Electronic Measurement

Signal Processing

Figure 1. Typical structure of absolute capacitance sensing system.

The physical sensor, which is the leftmost part in Fig. 1 is the element that serves as the capacitor to be sensed. We measure its capacitance to the external world. When a nger (or other conductive element) comes close enough to change this capacitance value, this change is measured and analyzed, and a nger location is calculated. Since the capacitance between the nger and the trace can be thought of (in zero-order approximation) as a parallel-plate capacitor,4 we have A , (1) CP arallelCapacitor = d where epsilon is the dielectric constant of the material between the plates, A is the area of the plates, and d is the distance between the plates. From (1) it follows that in order to have a large signal from an approaching nger, we need to have as much trace-area (A) exposed to the nger. Also, the smaller the distance (d) between the nger and the sensor, the larger the signal is. We will delve into this further when we talk about the measurement shortly, but it is worth remembering that we want to cover the sensing area with electrodes as much as we can, and keep all of them as close to the nger as we can. If we leave some area unused, we will be wasting signal that could have been captured. In addition, following up with the parallel-plate capacitor model for the interaction between the nger and the sensor, we can deduce that the signal (or change of capacitance on the sensor) due to a nger will be proportional to the overlapping area between the nger and the sensor. This is the regime where those two bodies constitute a parallel plate capacitor. We will use this fact extensively later on in the analysis and simulation parts.

Electronic measurement, the middle element in Fig. 1, is the part that measures the change in the physical property (capacitance), translates it to electronic signal, and usually convert it to digital-representation to be processed by the signal-processing block. In general, since capacitance describes the proportion ratio between voltage and charge, there are two dierent methods to measure it: imposing known voltage on the sensor and measuring the resulting charge on it; or alternatively, imposing known charge on the sensor and measuring the resulting voltage. Other methods, which measure the complex impedance of the sensor, follow essentially the same idea. We will not delve into the measurement process, as this is beyond the scope of this paper, however we want to emphasize that this is far from a trivial problem. One needs to deal with various issue, for example : noise (from LCD, other electrical components, etc); small signal-changes due to nger compared to background signal (the change due nger can be a few pico-Farads, whereas the background capacitance of the sensor can be around 50pF); need for high resolution in the reading in order to achieve good location accuracy (in the order of a few femto-Farads), and more. Most of the methods today use techniques involving modulation of the signal, averaging, and other noise-avoiding tricks to facilitate the measurement. Signal processing, the last part in Fig. 1, converts the signals from the sensor into nger coordinates, gesture recognition, and so on. This involves both traditional signal-processing, like signal conditioning, ltering (smoothing), and in addition algorithmic parts to analyze the data and deduce nger(s) location, tracking, and gestures interpretation. The nal result of this chain (sensor+measurement+analysis) should be robust to dierent usage modes (small/big nger, tap or draw, etc), manufacturing variability in the sensor, noise in the system, and at the same time give good results in terms of responsiveness, accuracy (local and global), and most importantly, not to produce disturbing instabilities that will frustrate the user. We will delve more into the requirements as we get to evaluate the dierent algorithms. The most common pattern for sensing absolute capacitance is described in Fig. 2, and is known as the diamond-pattern (for obvious reasons). It is composed of two layers, one called the columns-layer, and below it the rows-layer. This conguration of the electrodes complies with the goal of having maximum exposure to the nger, tiling the whole surface, and symmetrical for rows and columns. It does present the issue of the need for two layers, as otherwise the rows and columns will be short circuited. Most of our simulations will follow the use of this pattern, though they apply to other congurations as well.

R1 R2 Rows R3 R4

C1

C2

C3

C4

Columns

Figure 2. Illustration of the diamond pattern of electrodes, which is the most common pattern for absolute capacitance sensing.

In this pattern, one can consider the measurements as projections of the capacitance image onto rows and columns, where the rows and columns have the special concatenated diamond shape. This leads us to the

following problem formulation: Problem denition: Given the projections of capacitance-image over the sensor, nd the location of a nger on the sensor. Expanding a little more on this, we can elaborate: Expansion on problem denition: The resulting coordinates should be robust to variations in the measurements (noise, temperature, manufacturing), and variations in the nger model (of how nger looks like), where robust means accuracy and stability. In the general case, where the nger imprint on the capacitive image can be of any shape, and we may have multiple ngers altogether, this is of course an ill-posed problem. We have too few measurements to enable a unique correct solution. However, as we will show, there are cases where one can look for exact solution. In other cases, we will look for a stable solution, which is best according to some criteria. The rest of this work deals with addressing the above problem, and analyzing various solution options. One may wonder if this is a dicult problem: hasnt it been solved already many times? After all, there are many devices on the market using this technology already. The Moto-labs study5 of nger tracing accuracy on various phones shows the devices on the market have signicant errors when tracing straight lines. The study examined six dierent phones, and used a robotic-nger to trace a grid pattern of straight lines on the screen of each one. However, though the robotic movement is straight, the study found that the resulting lines reconstructed on the screen are often wiggly. We may conclude that the current generation of touchscreen phones has not solved the problem, and further investigation in the areas discussed in this paper are necessary.

3. PROJECTION APPROACH
Though the purpose of this paper is to investigate interpolation algorithms, it is useful to consider the inherent projection of the nger capacitance onto the row and column traces as a type of Radon transform, similar to what is studied in medical imaging.6 We know, by the Fourier slice theorem, that the Fourier transform at frequency of the projection at any angle is G(, ) = F ( sin , cos ), where F (u, v) is the 2-D Fourier transform of the function f (x, y). The nature of the projection as shown in Figure 2 involves sampling the averaged value over a square aperture, and summing the resulting samples. Space does not permit a derivation of the transform, but it is worth noting that the projection process may be modelled as follows. Suppose that nger capacitance is represented by a two-dimensional (2-D) function f (x, y) = f (x) with Fourier transform F (u), with x, u denoting 2-D vectors. The Fourier transform of the diamond-shaped aperture is denoted H(u); note that this will be a 2-D sinc function rotated 45-degrees. Then the aperture serves as a lowpass lter, and the observed signal is therefore Fo (u) = F (u)H(u). If is a hexagonal 2-D sampling lattice having fundamental unit area A(), with denoting the dual lattice, then the Fourier transform of the samples is Fs (u) = 1 A() Fo (u + k)
k

(2)

(3)

We can now derive the Fourier transform of the projection applying (2) to (3). The details are interesting; however, we leave those to future work because the purpose of this paper lies in investigating practical interpolation methods.

4. ANALYTIC SOLUTIONS
In this section we address two particular cases, which lend themselves to analytic analysis.

4.1 Wire Grid Sensor

In this section we will derive a solution for the case where the sensing electrodes are simple wires, as depicted in Fig. 3. In (a) the wire-grid is drawn, where we can assume each wire is much narrower than the pitch size, or the nger size. Moreover, the overlap between the wires reduces to such a small area, that we will also neglect that factor. Since the response of the sensor to the nger is proportional to the area overlap between the two, and since we assume thin wires with negligible overlap between themselves, the response to a round nger is reduced to being proportional to the length of the appropriate chord in the nger-circle. d R1 d R2 R3 R4 R1 R2 R3 R4 2 Rf

Rows

C1

C2

C3

C4

C1

C2

C3

C4

Columns (a) (b)

Figure 3. Wire-grid sensing electrodes pattern. (a) Basic conguration. (b) A round nger of size 2 Rf superimposed on the sensor from (a).

Now, let us assume that the nger radius is such that, unless at the edges of the sensor, it will always impact the reading on exactly two columns and two rows. This translates to the requirement that d > Rf > d/2, where d is the spacing between traces. Fig. 4 describes such a scenario, depicting only the columns sensors. In this case, only columns 1 and 2 will have readings, and we will denote those values as c1 and c2 , respectively. Since we assume the response is proportional to the area covered by the nger, we immediately note that c1 and c2 are nothing but the respective chords of a nger-circle. With the help of some basic geometry, and the notations in Fig. 4, we can write the following 3 equations: x1 x2 x1 + x2 = = = d. c1 2 c2 Rf 2 2 Rf 2
2

(4)

,
2

(5) (6) (7)

These comprise 3 independent equations, with 3 unknowns: x1 , x2 , and Rf . From substituting the rst two in the third one, we get one (non-linear) equation with only one unknown Rf : d= Rf 2 c1 2
2

Rf 2

c2 2

(8)

d x1 x2

Rf

C1

C2

C3

C4

Figure 4. Notations for analytic computation.

There are a few things to note about this result: Analytic solution - The above is an exact solution for determining the x-coordinate and Rf , the radius of the nger. Considering the fact we have three unknowns ((x, y) coordinates of the center of the nger, and Rf the radius), and 4 measurements (2 on rows and 2 on columns), it is not surprising we are able to nd a unique analytic solution. Separation of variables - It is worth noting that the X and Y coordinates can be computed independently. However, keeping in mind we do have 4-measurements for only 3 unknowns, we can exploit this to solve the whole system of equations together in order to gain robustness to noise. Still, it is a closed solution, but is considering the possibility the noise, and thus is more robust. Stating it yet dierently, the fact that Rf is the same for both axes, creates the coupling that can be exploited to overcome noise. Solving in real-time: LUT - Since we can write the equation in terms of Rf = Rf (c1 , c2 ), we can use a two-dimensional look-up-table to solve the equation in real-time. In this very symmetric sensor case, we can use the same table for the X and Y axes. The size of the LUT depends on the resolution we want in the X/Y directions, and the resolution of our measurements.

4.2 Analytic solution for square nger

We follow the method used above, but for the case of a square nger. Note that the diamond grid in Figure 2 is the same as a square grid rotated 45 degrees. Suppose that we now have exactly one square nger, aligned to the grid. Assume that the width of the square, W , satises a constraint similar to (4): d > W > d/2. (9)

Here again d is the spacing between the electrodes. With respect to Figure 2, the constraint means that we obtain a non-zero reading on at least one trace Ri and one trace Cj , but not more than two of either the row or the column readings are nonzero. When it produces a nonzero reading on one and only one Ri and Cj , then the nger ts within the basic block of Figure 5. Denoting as before the readings on each trace R or C by lowercase r or c, we see that, within this block, the nger position is not uniquely determined by the ri , cj values. We now show that there are in general four symmetries, and that the contours of equal readings are hyperbolic in shape.

uy

Ri

Cj
Figure 5. A basic block of 2 2 diamond sensors is shown, with the horizontally connected row traces and vertical column traces indicated. The square nger, assumed to be aligned to the grid, has width W , which is larger than a single diamond, but small enough to t within the basic block.

From Figure 5 the reading on the hatched electrode for a square nger having position (ux , uy ) is ri = ux uy + (W ux )(W uy ) = 2uxuy + W 2 (ux + uy )W. (10)

Neglecting the constant term W 2 on the right, it can be seen that the contour lines of the 2uxuy term are hyperbolic in shape, and that the contour lines of the (ux + uy )W term provide a linear correction. We can solve (10) for uy as follows: ri (W ux )W uy = . (11) 2ux W We obtain the same equation for the column reading on the plain electrodes in Figure 5, namely that cj = W (ux + uy ) 2ux uy (12)

Clearly, ri + cj = W 2 , an equation which when combined with (10) and (12) gives us three equations to solve for ux , uy , and W . Note that by the algebraically-symmetric form in (10) and (12), if (ux , uy ) is a solution, then so is (uy , ux ), as well as (W ux , W uy ) and (W uy , W ux ); hence there is a four-fold symmetry to the solution space. Figure 6 illustrates the symmetries by showing the contour lines of ri , cj ux , uy space. on The contour lines were obtained by simulation, with W = 75, and the trace spacing is d = 100 2. The range of coordinates ux , uy that t within a basic block is therefore 25.

5. SUBPIXEL INTERPOLATION
Suppose that a nger gives readings on more than one electrode ri and more than one electrode cj . The problem of nding the position to a resolution smaller than d, the spacing of the diamond-shaped blocks in Figure 1, is formally similar to the problem of nding the peak location of a continuous function to subpixel level. We describe briey several popular algorithms for subpixel peak estimation.7 Below, assume that ri is the maximum value of the row trace readings, ri1 is the row below the maximum, ri+1 is above, and similarly for the column readings. Table 1 summarizes four dierent algorithms for estimating subpixel location of the nger from the three samples (ri1 , ri , ri+1 ). Similar equations hold for the column readings (ci1 , ci , ci+1 ). To gain insight into the estimators used in the table, let us consider, for example, the estimator using the center of mass method. From the table, the equation is = ri+1 ri1 . ri1 + ri + ri+1 (13)

ri contours 25 20 15 r 10 5 5 10 15 ux 20 25 25 20 15 10 5 5

cj contours

uy

uy

10

15 ux

20

25

Figure 6. Illustration of four-fold symmetry in contour lines for row (column) readings for a square nger that ts in a basic block. The arrows indicate the direction in which r or c increase, respectively; note, for example, that r increases if ux , uy both increase, so moving diagonally upward in this gure is equivalent to moving the nger horizontally in Fig. 5. This gure, and others to follow, is best viewed in color.

Estimator Gaussian Center of mass Linear Parabolic

Equation
ln ri1 ln ri+1 1 2 ln ri1 2 ln ri +ln ri+1 ri+1 ri1 ri1 +ri +ri+1 ri+1 ri1 1 2 ri min{ri1 ,ri+1 } ri1 ri+1 1 2 ri1 2ri +ri+1

Table 1. Subpixel nger position estimators

The upper bound ri+1 is reached if ri1 = 0, which means ri+1 1 since ri+1 + ri = 1 and ri ri+1 by the 2 1 1 assumption that its the peak location. Hence 2 and we can show similarly that 2 . For any value of , the estimated location of the peak is i+, where i is the integer index of the maximum row reading, and serves as the subpixel oset. For the other estimators, some insight may be gained by noting that the parabolic estimator is the solution of a second-order Taylor series expansion of the nger capacitance; the Gaussian estimator is similar, up to the log function, because the exponent of a Gaussian distribution is parabolic. Similarly, the linear estimator assumes that the shape of function is a straight line on both sides of the peak; see Fisher and Naidu7 for details.

The center of mass estimator may be derived by assuming that (ri1 , ri , ri+1 ) form a discrete probability distribution located on the indices (1, 0, +1), and solving for the mean of the distribution. Note that we assume that ri is the peak value of the distribution; this implies || 1 , as we now show. In (13), we may set the 2 denominator ri1 + ri + ri+1 to 1, by dividing both top and bottom by the sum. The numerator has upper and lower bounds ri+1 ri+1 ri1 ri1 (14)

6. SIMULATIONS AND DISCUSSION

A Matlab simulation was used to compare the various methods discussed under dierent conditions. The diamond pattern used is described in Fig. 7. In (a) we describe a single row of diamond pattern. There are 16 rows in our simulated panel. In (b) a single column is described. There are 12 columns in our simulated panel. In (c) we combine rows and columns to create a full sensing panel. Each row and column are colored dierently, and thus we have 12 + 16 = 28 dierent colors in the picture. Typical dimensions for the row and column spacing (and size) is about 5mm. This has to agree with the expected nger size, which is around 7mm to 15mm in diameter. Keep in mind the size is the one captured by the capacitive sensing electrodes, and thus takes into account fringe-elds etc. In our simulation, the basic pitch, namely the distance between the center of any two rows (or columns) was discretized into 100 cells. Therefore, assuming a 5mm pitch, a pixel in the simulation is of size 50m. We used a round nger for the simulation, and added a prole to its capacitive characteristics. Fig. 8 combines dierent views of the simulated nger. We used a Gaussian prole truncated to t into a diameter of 3 columns and rows. Since the pitch is 5mm, it means that the diameter of the nger inuence is 15mm. This takes into account the thickness of the cover-glass and fringe-eld eects. Once we have a simulated sensor, and a simulated nger, we can evaluate the interaction between those. We do this by placing the nger with a center on a known location of the sensor, as described in Fig. 9(a). We can then compute the overlap between the nger (as weighted by the prole) and each of the sensors (rows or columns). This will give us the reading of the rows and columns, which are shown in Fig. 9(b-c). Using these proles as a starting point, we can then compute a location according to various algorithms, and compare results. Keep in mind we also know the exact location of the center of the nger, which is where we simulated it in the rst place. Thus, we can also calculate error measures. Taking it one step further, we can then simulate the nger performing a trajectory over the sensor. We simply put it each time in a new location, and for each location compute the proles, and calculate simulated location. This will give us indication of periodic errors, and errors which depend on the location of the nger relative to the grid. The results of such a process are described in Fig. 10. One can clearly see the periodic nature of the error, which is the result of the interaction between the nger trajectory and the underlying periodic diamond pattern. In addition, it is evident that some algorithms perform better than others, in terms of maximum-error and in terms of mean-error. We then introduce multiplicative noise component to the raw prole data, and the results are given in Fig. 11. The amplitude of the noise is varied, and indeed one can see the dierent impacts of noise on the various algorithms. Some algorithms are much more sensitive to noise. One can compare the algorithms under dierent kind of noises (additive, white, correlated, and so on). In addition, dierent nger sizes, and shapes, will also eect the dierent algorithms dierently. Results for those cases will be shared in the presentation.

7. SUMMARY AND CONCLUSION

In this work we discussed the most common method of projected capacitive sensing, that of absolute capacitive sensing. We analyzed the problems involved in determining nger location from the most-common type of sensors, diamond-shaped sensors. We formulated the problem as a reconstruction from projections, and considered two simple case. In the rst case, that of circular nger, we derive analytically the solution for nger location. We repeat the analytical approach for a second case of interest, that of a square nger. We describe interpolation algorithms, and evaluate their performance. This work just touched the tip of the iceberg, and there are many more issues to explore. Non-separable algorithms, plausibly using an LUT implementation, might give us a hint at the optimal solution. Specic algorithms which are optimal with regard to insensitivity to nger size or shape (within limits), are also of paramount importance. In addition, the projections approach, using inverse transform of the projections, presents a fertile eld for investigation.

2 4 6 8 10 12 14
2 4 6 8 10 12

16

(a)
Full panel: 28 traces 2 4 20 6 8 10 10 12 14 16 5 15 25

(b)

10

12

(c)
Figure 7. Panel diamond pattern: (a) Single row of diamond pattern.(b) Single column of diamond pattern. (c) Combining rows and columns to create a full sensing panel.

Finger cross section 1

Finger

0.9

0.8

0.7

0.6

0.5

0.4

(a)

(b)

Figure 8. Simulated nger prole used in the simulations: (a) Color depicting magnitude. (b) Cross section of prole.

Array with finger area superimposed

2 4 6 8 10 12 14 16

10

12

(a)
Traces response: rows 10000 9000 8000 7000 6000 5000 4000 3000 2000 1000 0 0 2 4 6 8 10 12 14 16 0 1 2 3 4 5 6 7 8 9 10 11 12 4000 6000 10000 12000 Traces response: cols

8000

2000

(b)

(c)

Figure 9. (a) Finger superimposed on the Diamond pattern. (b) Response on the rows. (c) Response on the columns.

Blue exact route Red COM (max/mean error = 0.700224 // 0.417365 [mm]) Green Linear = 0.519937 // 0.317378 [mm]) Black Gaussian = 0.385795 // 0.209413 [mm]) Cyan Parabolic = 0.252979 // 0.148013 [mm]) 1600 1400 1200 1040 1000 800 600 400 940 200 920 100 200 300 400 500 600 700 800 900 1000 1100 1200 900 450 1020 1000 980 960 1100 1080 1060

Blue exact route Red COM (max/mean error = 0.700224 // 0.417365 [mm]) Green Linear = 0.519937 // 0.317378 [mm]) Black Gaussian = 0.385795 // 0.209413 [mm]) Cyan Parabolic = 0.252979 // 0.148013 [mm])

500

550

600

650

(a)

(b)

Figure 10. (a) Simulating the nger at a dierent location, traveling along a straight line, and for each location calculating the response on the traces (rows and columns), and from that the estimated computed location. The exact travel location is depicted as a Blue line, where the others depict dierent interpolation algorithms (see title of picture for more details). (b) Zooming in on part of the graph in (a).

Blue exact route Red COM (max/mean error = 0.906636 // 0.432628 [mm]) Green Linear = 0.676228 // 0.332286 [mm]) Black Gaussian = 0.551196 // 0.236180 [mm]) Cyan Parabolic = 0.556204 // 0.235776 [mm]) 1100 1080 1060 1040 1020 1000 980 960 940 920 900 450 500 550 600 650 1100 1080 1060 1040 1020 1000 980 960 940 920 900 450

Blue exact route Red COM (max/mean error = 1.237692 // 0.559910 [mm]) Green Linear = 1.264948 // 0.532050 [mm]) Black Gaussian = 1.149167 // 0.446962 [mm]) Cyan Parabolic = 1.490748 // 0.504834 [mm])

500

550

600

650

(a)

(b)

Figure 11. Same trajectory as in the previous image, only this time a multiplicative noise is added to the measurement. (a) Noise is uniformly distributed with amplitude of 0.1. (b) Noise is uniformly distributed with 0.3 amplitude.

REFERENCES
[1] M. R. Cutkosky, R. D. Howe, W. R. P., Force and tactile sensors, in [Springer handbook of robotics], Siciliano, B. and Khatib, O., eds., Springer-Verlag, Berlin (2007). [2] Walker, J., Touch technologies tutorial. http://www.walkermobile.com/PublishedMaterial.htm (2010). [Online: accessed 8 Dec 2010]. [3] Baxter, L., [Capacitive Sensors: Design and Applications], IEEE-Wiley (1996). [4] Halliday D, Resnick R, W. J., [Fundamentals of Physics], Wiley (2004). [5] MotoLabs, Diy touch screen analysis. vimeo.com/motodevelopment (2010). [Online: accessed 10 Jan 2010]. [6] Herman, G. T., [Fundamentals of computerized tomography: Image reconstruction from projection ], Springer, 2nd ed. (2009). [7] Fisher, R. B. and Naidu, D. K., A comparison of algorithms for subpixel peak detection, in [in Image Technology, Advances in Image Processing, Multimedia and Machine Vision ], 385404, Springer-Verlag (1996).