See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4340049

Line tracking algorithm for scribbled drawings

Conference Paper · April 2008

DOI: 10.1109/ISCCSP.2008.4537287 · Source: IEEE Xplore

CITATIONS READS

9 1,175

4 authors:

Alexandra Bonnici Kenneth P Camilleri

University of Malta University of Malta
52 PUBLICATIONS 252 CITATIONS 180 PUBLICATIONS 3,491 CITATIONS

SEE PROFILE SEE PROFILE

Simon G Fabri Jonathan Borg

University of Malta University of Malta
115 PUBLICATIONS 2,906 CITATIONS 171 PUBLICATIONS 5,151 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Alexandra Bonnici on 04 November 2014.

The user has requested enhancement of the downloaded file.

 Line Tracking Algorithm for Scribbled Drawings
Alexandra Bartolo∗ , Kenneth P. Camilleri∗ , Simon G. Fabri∗ and Jonathan C. Borg†
∗ Departmentof Systems and Control Engineering
University of Malta
Email: [abbart, kpcami, sgfabr]@eng.um.edu.mt
† Department of Industrial and Manufacturing Engineering

University of Malta
Email: [email protected]

Abstract—This paper describes a line tracking algorithm that and Section VI presents the conclusions that can be derived
may be used to extract lines from paper based scribbles. The from this study.
proposed algorithm improves the performance of existing sparse-
pixel line tracking techniques that are used in vectorization by II. E XTRACTING L INES FROM S CRIBBLED D RAWINGS
introducing perceptual saliency and Kalman filtering concepts to
the line tracking. Furthermore, an adaptive sampling size is used Paper-based line drawings cannot be used as direct input to
such that it is possible to adjust the size of the tracking step to CAD systems since the digital representation of the scribbled
reflect the stroke curvature. drawings is saved in a raster format, whereas the CAD tools
typically require graphics represented in vector format. Thus,
I. I NTRODUCTION
in order to edit the paper-based drawing using the CAD tools,
Scribbling is a fast and efficient method through which the designer’s line strokes must be extracted from the raster
mental ideas are formalized. Although digital tablet technology image and represented with appropriate mathematical models
is developing at a fast rate, it is unlikely that digital tablets whose parameters may be estimated. This may be carried out
will completely replace paper-based scribbling, particularly in through vectorization, which may be broadly described as a
shape and form design applications where designers tend to three-step process consisting of line location, line fitting and
prefer paper-based scribbling to digital scribbling [1]. Scrib- line beautification.
bles however, cannot be integrated directly with CAD tools Vectorization techniques have been traditionally developed
which are used in subsequent design stages. Sketch based for applications involving neat drawings, such as engineering
interfaces that integrate the scribbles with CAD tools are there- drawings [5], architectural drawings [6] or cadastral maps [7].
fore desirable so that the flexibility of paper-based drawings Since such lines are rarely of a single pixel thickness, a line lo-
can be augmented with the 3D virtual CAD modeling, giving cation procedure is required to eliminate redundant pixels from
designers the opportunity to invest more time on design rather the line drawing, retaining only those which are necessary to
than interfacing with the CAD tools. maintain the shape and connectivity of the object [8]. Thus,
Research in this area may be broadly divided into two the line drawing is represented by a thin sequence of pixels
categories, namely the creation of algorithms that interpret which may be referred to as the medial axis of the drawing [9].
paper-based drawings and the creation of interfaces that make However, although this representation retains only the pixels
digital sketching as close as possible to paper-based sketch- necessary to preserve the topology of the line drawing, the
ing. Unfortunately, paper-based interpretation algorithms are representation is still in a raster format and further processing
mostly restricted to neat, technical drawings and do not han- is required in order to obtain the vector representation as
dle sloppy or scribbled drawings, whereas online interfaces, typically required by CAD tools. This is carried out through
while allowing the interpretation of scribbled drawings, require the line fitting step which splits the connected sequence of
human interaction and do not allow designers to use their pixels such that each pixel group may be approximated by
preferred drawing medium. equations of straight lines or curves of known parameters [8].
This paper gives the details of a line tracking algorithm Most interpretation systems apply a subsequent, final line
which forms part of a scribble simplification procedure de- beautification step to this vector data before using the CAD
scribed in [2]. This scribble simplification replaces traditional tool [8]. This beautification is necessary to correct errors that
line location techniques such that when used in conjunction may be due to the freehand nature of the drawing. Thus each
with an annotation-based sketching language described in [3], vector is compared to its neighbouring vectors in order to
allows designers to create 3D CAD models from paper-based identify relations such as collinearity or parallelism, hence
scribbles. introducing a degree of contextual knowledge to the vector
The rest of the paper is divided as follows: Section II gives a data [8].
brief review of existing line extraction and vectorization algo- The line location procedure is the most crucial step in the
rithms, Section III presents our proposed solution; evaluation vectorization process since this determines the accuracy of
of the proposed solution and results are presented in Section V; the resulting vectors with respect to the line drawing. For
this reason, the main difference between various vectorization
algorithms is found in the line location step, with the goal
of each algorithm being to find lines accurately with minimal
computational expense. Vectorization techniques can therefore
be divided into six main categories according to the techniques
used to locate lines namely, Skeleton based methods, Con-
tour based methods, Skeleton-Contour based methods, Hough
based methods, Run-graph based methods and Sparse-pixel (a) (b)
based methods [10].
Fig. 1. The piecewise linear path propagation is accurate in straight
The scribbled drawings have a mixture of linear and curved line segments such but may cause the path line to deviate from the medial
stroke segments, such that the line location algorithm should axis points in curved segments as shown in (b). To solve this, scan lines
be capable of extracting linear and curved strokes without perpendicular to the line segment are taken such that the tracking point is
placed in the middle of the line profile as shown in (c)
causing unnecessary segmentation of the strokes or require
additional processing to locate the curved strokes. It is also
necessary that the line location represents the strokes using
the required orientation estimates. Using these orientation
the minimal number of points without under-sampling the
estimates, the co-ordinates of the tracking point at an tracking
strokes. Finally, unlike neat drawings, the scribbled drawing
instant k + 1 may be determined from the position of the track
is not necessarily an exact representation of the designer’s
point at a tracking instant k using (1), where θk is the line
intent, such that perceptual grouping of strokes is necessary
orientation at point (xk , yk ) and D is the distance between
to identify the most likely object shape.
two consecutive line points.
Existing line location techniques do not meet these criteria.
In Skeleton and Contour based line location methods, the rep-
resentation contains all the points on the medial axis. Although xk+1 = xk + D cos(θk )
Run-based and Sparse-pixel based algorithms represent the yk+1 = yk + D sin(θk ) (1)
medial lines by a sample of points, the sampling rate is a
user-defined parameter such that the designer must identify This piecewise linear tracking works well for straight line
a suitable sampling rate for the drawings. Furthermore, Run- segments but in the case of curved line segments, determining
based algorithms and some Sparse-pixel algorithms are more the next point using (1) alone may cause the tracking points to
suitable for straight lines, segmenting curves unnecessarily. offset from the line’s medial points as shown in Fig. 1(a). This
Other Sparse-pixel algorithms and Hough-based algorithms offset becomes more evident when the distance D between
require a different line location process to locate the curves. consecutive points is increased. In order to adjust for this
These algorithms assume that the curves may be represented offset, the orientation of the line at the new point is determined
as circular arcs, however, this restricts the objects that may such that a scan line perpendicular to the line segment can be
be represented, particularly, free-form object profiles which obtained. As shown in Fig. 1(b), this scan line gives a measure
may be more suitably represented by splines. Furthermore, of the line’s profile such that the midpoint (mx , my ) of the
none of the vectorization algorithms described in the literature line may be determined. This point can therefore be considered
combines perceptual selection of strokes with the vectorization as a better estimate for the point (xk+1 , yk+1 ) determined
process. through (1). The tracking procedure is repeated until either an
end of line is reached or the tracking path reaches a segment
III. T HE P ROPOSED L INE T RACKING A LGORITHM region that has been previously tracked.
This work proposes the use of a sparse pixel tracking The tracking method described above is very similar to the
approach, similar to the Sparse Pixel Vectorization (SPV) tracking algorithm described in SPV [5]. However, in SPV
algorithm described in [5]. This approach has been selected tracking, the initial tracking step is carried out by moving
as it allows the introduction of the Gestalt laws of perceptual the tracking point by a distance D either in a horizontal or
grouping into the line tracking procedure and this, together in a vertical direction since the SPV tracking process has
with other modifications described below, enhances the per- no prior knowledge of the line direction θ. This also affects
formance of the SPV algorithm. the construction of the scan-line which may be constructed
The tracking algorithm is performed after processing the perpendicular to the line stroke, rather than using a vertical
scribble by Gabor-filter based algorithm to group scribble line or horizontal scan of the line. Thus, the midpoint (mx , my )
strokes together therefore, information on the orientation of obtained form the line profile is more reliable when evaluated
each line passing through each pixel is known [2]. These using the proposed method than when using the SPV method.
orientation estimates for the basis of the perceptual selection Using the line direction θ to evaluate the position of the next
of the line tracking direction at junction points. Since the point has another advantage, namely that changes in the line
proposed line tracking algorithm is algorithmically indepen- direction are directly reflected as changes in the value of θ,
dent of the Gabor filter grouping algorithm, any other line such that curved paths may be tracked in a piece-wise linear
group orientation estimation method can be used to obtain manner with no adjustments to the tracking procedure. In
 at each tracking step. For straight line regions, the orientation
at each tracking step is constant such that when (1) is used to
determine the initial position of the next tracking step, this will
be located within the stroke boundaries of the line segment.
However, in curved strokes, (1) retains its validity for short
tracking steps only such that if the tracking step is too large,
(a) (b)
the initial position is located outside of the stroke boundaries.
Furthermore, a change in line orientation is measured at each
Fig. 2. The tracking step D should reflect the properties of the line segments tracking point, giving further indication of the curvature of the
being tracked. (a) Using a constant tracking step size would result in poor strokes.
track points in regions of high curvature. (b) A dynamic tracking step that
adjusts according to the curvature allows finer tracking in regions of high For a given line path, the tracking step is too large when (1)
curvature and a coarser tracking step in straight line regions. propagate the tracking point to a new point on the image
background. When this occurs, the tracking step is halved in
an attempt to find an initial tracking point that lies within the
SPV however, a change in the line curvature of more than stroke boundaries. If this is found then tracking resumes, using
π
4 requires a change in the tracking direction from horizontal the new tracking step. If however, the new initial point is still
to vertical or vice versa. Such a change in direction is treated beyond the stroke boundaries, the tracking step is again halved
as a junction of two line segments such that the curved path and the process repeated until a minimum tracking step of one
cannot be extracted as a single entity, but is broken into smaller pixel is reached, in which case the tracking is terminated.
fragments. On the other hand, a straight line region is detected when
the tracked points are on the stroke foreground and the line
A. Adaptive Tracking Step Size orientation of the track points is constant. When this happens,
Similar to the SPV algorithm, a tracking step size D must the tracking step is incremented until either a maximum track
be initialized in order to perform the line tracking. This track step is reached or the stroke segments are no longer linear.
step determines the sampling rate of the line path and must Thus as shown in Fig. 2(b), it is possible to change the tracking
therefore be selected with care. Small values of D would step size such that the tracking uses a fine or coarse step size
result in a large number of sample points, increasing the data according to the characteristics of the line.
points extracted from the drawing and hence the computational
time of the path following. On the other hand, large values B. Treatment of Junction Points
may result in under-sampling of curved paths, as shown in The stroke orientations obtained through the Gabor filter
Fig. 2(a). Therefore, it is necessary to determine a tracking provide to the proposed line tracking algorithm a distinct
step that retains only a subset of all the pixels forming the advantage over the SPV algorithm in the treatment of junction
line path, while at the same time keeping a sampling step regions. Using SPV, the line tracking at junctions is terminated
that retains sufficient points to represent curves accurately. unless it is possible to continue tracking in the same tracking
Since the drawing may contain a range of segments with direction. In the proposed tracking algorithm however, path
varying degrees of curvature, the tracking procedure ideally tracking may be pursued in any of the orientation estimates,
changes the track step adaptively such that large tracking corresponding to each line stroke present at the junction
steps are used in linear regions which may be defined by region. It is therefore necessary to select an orientation that
fewer tracking points, while using a finer tracking step size for allows the tracking algorithm to follow paths that are perceived
curved regions whose representation requires a larger number as more salient than others. This motivates the use of the
of sample points. Gestalt laws of perceptual grouping which state that long,
In SPV, Liu and Dori change the value of the tracking step smooth continuous paths are preferred to others having a
D when the tracking point encounters a junction region, in larger number of curvature changes while at the same time
which case the tracking step is halved in an attempt to locate giving preference to closed objects over meandering paths.
the exact start of the junction region. If the start of the junction This implies that although smoothly continuous tracking is
is not located, the tracking step size is again halved and this often a good tracking guide, the tracking procedure should
is repeated until either the tracking step size is reduced to not adhere strictly to this rule since a less smooth path can at
one pixel width, in which case the tracking is stopped, or times prove to give a better tracking solution. For example, at
the junction boundary is located. If the junction boundary is P1 in Fig. 3, smooth continuation would be preferred as this
located, then the tracking step is reset to its original value allows the tracking to retain the outer contour of the object.
in an attempt to proceed with the path tracking. Thus, SPV However, at P2, the smooth continuation would not support
only reduces the step size if a junction region is encountered, further tracking such that a less smooth tracking path would
otherwise, the tracking step is kept at a constant value. be preferred.
In order to obtain an adaptive tracking step, the orientation In this work we introduce the Gestalt laws to the tracking
of the strokes as well as the location of the tracking point, prior process in order to select the most salient orientation at
to its being moved to the centre of the scan line, are monitored junction regions where two or more line orientation estimates
 smoothness of the stroke boundaries. The reliability of the
initial position (xk , yk ) depends on the orientation resolution
of the quadrature Gabor filter, and in the case of curved
strokes, on the tracking step size used. The second position
(mx , my ) is generally more accurate than the initial position
since by locating the midpoint of the scan lines, the second
point is adjusted such that it is closer to the medial axis of
the line strokes. However, at junctions, where the accuracy of
the scan lines is compromised, such that the error in the line
boundaries may be greater than the orientation resolution of
Fig. 3. The smooth continuation criterion allows the tracking to follow the the Gabor filter, the initial tracking point is likely to be more
outer car strokes at P1. However, smooth continuation is not always desired accurate than the scan-line midpoints. Thus it is desirable to
since in instances such as P2 a change in the path curvature is required in have a system that determines which of the two tracking points
order to maintain tracking.
is more reliable, given the noise present in the line boundary
at that instant and the noise due to the Gabor filter resolution.
are obtained. The measure of saliency used has two parts, the The Kalman filter framework provides such a system by
first evaluates the smoothness of the path while the second using the noise co-variance to determine a weight matrix
evaluates the certainty with which the line parameters are that can be used to obtain a weighted average of the two
known for each line at the junction. Since the line segments are tracking measurements. Although the Kalman filter is usu-
only known up to the current tracking point, it is necessary ally associated with dynamic systems, the Kalman filter has
to make a preliminary track step in each of the directions been applied to non-dynamic applications in order to track
indicated by the dominant orientations obtained from the roads [12], characters in text documents [13] and lines in
Gabor filter responses. Each of the dominant orientations will document images [14].
result in a preliminary track point (px , py ) which will be used
to evaluate the smoothness of the path on which it lies as well IV. K ALMAN F ILTER T RACKING
as the confidence of its neighbourhood. The discrete time Kalman filter assumes that a system is
The smoothness function used by [11], given by (2) is governed by a process modeled by the linear stochastic model
adopted, where α is the angle that that each preliminary given by (4) for which a measurement z may be related to the
tracking point (px , py ) makes with the tacked path generated system’s state vector x using (5) [15].
up till the current sample point.

!
2α tan α2 xk = Axk−1 + Bu + wk−1 (4)
f = exp (2)
D zk = Hxk + vk (5)

However, the smoothness function does not determine if where w and v are random variables representing the process
it is possible to maintain tracking in the selected direction, and measurement noise respectively which are assumed to be
hence the need of a second saliency measure which determine white Gaussian noise processes having zero mean and co-
whether tracking continuation is possible. This may be ob- variance matrices Q and R respectively. Matrix A relates the
tained by observing the pixel intensity I(px , py ) at each of state at the previous time instant k−1 to the state at the current
the preliminary track points (px , py ), where I represents the time instant k, matrix B relates the optional control input u to
response obtained by a stroke-grouping algorithm such as the the state x and matrix h relates the state x to the measurement
Gabor filter response. The total saliency s of the path can z.
therefore be defined by (3). The Kalman filter estimates the system state in two steps,
referred to as the time update step and the measurement update
s = I(px , py )f (3) step such that feedback obtained from the measurement z is
used to obtain better estimates of the state x. In the time update
Thus the saliency of the candidate line paths at the junction step, a prediction of the system state and error covariance is
region depends on the smoothness of the path with respect to made using the knowledge gained about the process model,
the current path as well as the possibility to extend the tracking the system states and the error covariance up till just prior to
beyond one tracking step. the current time instant. These predictions are referred to as
The tracking algorithm described above determines two the a priori state estimate and the a priori error covariance
positions of the track point at each tracking step k. The first, and are denoted as xk|k−1 and Pk|k−1 respectively. They are
initial track point is obtained through the line propagation de- given in (6).
fined by (1) and is therefore dependent on the line orientation
θk only. The second position is determined from the scan line Xk|k−1 = AXk−1|k−1 + BUk−1|k−1
of the strokes and is therefore dependent on the accuracy and Pk|k−1 = APk−1|k−1 AT + Q (6)
 The a priori state estimates and error covariance are then present also in the Kalman filter implementations described in
updated in the measurement update step to obtain state esti- [12]–[14]. However, unlike these methods, the Kalman filter
mates that take into consideration the most recent measure- is not applied directly to the image but to a set of data points
ments obtained from the system at the current time instant. that approximate the medial axis and which are extracted from
These updated estimates are referred to as the a posteriori the drawing prior to its being processed by the Kalman filter.
state and error covariance and are denoted as xk|k and Pk|k Thus, in this implementation, the causal nature of the Kalman
respectively. The measurement update step, defined by Equa- filter may be removed by using the Smoothing Kalman filter.
tions 7 compares the actual measurement zk to the ‘predicted’ This estimates the smoothed distributions conditioned to all
measurement obtained when the measurement model H is measurement data z1:N where N is the total number of mea-
applied to the a priori state estimates xk|k−1 surements. The smoothed distributions may be calculated from
the Kalman filter results by using the backward recursions
k = n, n − 1, · · · 1 on (9) [16].
Kk = Pk|k−1 H T (HPk|k−1 H T + R)−1
Xk|k = Xk|k−1 + Kk (zk − HXk|k−1 ) (7) −1
Jk = Pk|k+1 A(Pk+1|k )
Pk|k = (I − Kk H)Pk|k−1
Xk|N = Xk|k+1 + Jk (Xk+1|N − Xk+1|k ) (9)
A gain matrix K, which is estimated to minimize the a
Pk|N = Pk|k + Jk (Pk+1|N − Pk+1|k )JkT
posterior error covariance, is used to weight the difference
between the two measurements and hence update the a priori V. E VALUATION AND R ESULTS
state estimate xk|k−1 to obtain the a posteriori state estimate In order to determine the performance of the proposed line-
xk|k [15]. location algorithm, the algorithm was compared to the line
A. Application of the Kalman filter to line tracking location obtained by the SPV algorithm, using a number of
test images containing a mixture of straight line strokes and
The scribbled line stroke may be modeled by a pen moving curved strokes such as those shown in Fig. 4(a) and Fig. 4(c).
along a trajectory in a piecewise linear manner such that the The performance of the two algorithms was measured using
position of the pen defined by its (x, y) co-ordinates at any the Pixel Recovery Index (PRI) defined in [17] which measures
instant k may be written as the location of the line strokes extracted by the line location
algorithm with respect to the ground truth data. The proposed
xk+1 = xk + ∆xk algorithm obtained a PRI of 87% whereas the SPV algorithm
yk+1 = yk + ∆yk (8) obtained a PRI of 82%, hence the lines extracted from the
drawing using the proposed line tracking algorithm are better
where ∆xk = D cos θk ∆yk = D sin θk and θk is the line to the ground truth drawings than those extracted by the SPV
orientation obtained from the orientation estimates at the co- algorithm. Furthermore, the proposed line tracking algorithm
ordinates (xk , yk ). The medial point obtained from the line reduces the computational time of the SPV algorithm by
profile taken perpendicular to the line path may be considered 56%, hence reducing the time to obtain the stroke lines while
as the system measurements, obtaining an alternative measure improving the line quality.
for the line position of the tracking point. The performance of the Kalman filter was also compared to
In this way, the scribble may be considered as a process smoothing obtained by simple moving spatial average filters.
modeled by (8) having a state vector x = [x, y]T and input This was carried out by adding random positional noise having
u = [∆x, ∆y]T . From the stroke-grouping response image I, a uniform distribution in the range [−5, 5] pixels to the medial
measurements giving the position of the medial points may lines, in order to simulate the contour noise. The error between
be obtained. By considering these measurements as the output the smoothed lines and the ground-truth lines can therefore be
of noisy sensors, the Kalman filter may be used to estimate measured, obtaining a measure of the smoothing performance
the system states, and obtain better estimates of the medial of the Kalman filter in comparison to the moving average filter.
axis. By comparing between (8) with (4) matrices A and B The moving average filter requires the specification of a
reduce to identity matrices of size 2 × 2. Furthermore, the window size which determines the number of medial points
measurements of the medial axis points are identical to the over which smoothing is to take place. Larger window sizes
system states hence the measurement model H in (5) is also are desirable in straight line regions as these would allow
a 2 × 2 identity matrix. more smoothing of the medial lines. However, large window
Since the Kalman filter performs the state estimates using sizes would shift the medial lines to the centre of curvature of
only the current and previous states and measurements, the curved paths. Thus in moving average filtering, a compromise
filter is causal. This implies that initializing the filter from between smoothing and displacement of medial lines must
different starting points and using different directions, the be reached. For the medial lines used in this test, which
Kalman filter would result in different state estimates. This is consisted of medial lines extracted from scribbles similar to
undesirable since the medial axis of the line drawing should those shown in Fig. 4, a window size of 7 pixels was found to
not be affected by the tracking direction. This problem is be more suitable, maximizing the error reduction to 37.85%.
 algorithms. The line location requires subsequent processing to
link the extracted lines forming completely connected objects.
This is typically carried out through the beautification step.
In future, we propose to introduce further perceptual grouping
techniques to this beautification step in order to select the most
perceptually salient link formation.
(a) (b)
R EFERENCES
[1] J. A. Landay and B. A. Myers, “Sketching Interfaces: Toward More
Human Interface Design,” IEEE Computer, vol. 34, no. 3, pp. 56–64,
March 2001.
[2] A. Bartolo, K. P. Camilleri, S. G. Fabri, J. C. Borg, and P. J. Farrugia,
“Scribbles to vectors: Preparation of scribble drawings for cad inter-
pretation,” in Eurographics Workshop on Sketch-Based Interfaces and
Modelling, M. V. d. P. Eric Saund, Ed., 2007.
[3] A. Bartolo, K. P. Camilleri, P. J. Farrugia, and J. C. Borg, “A new
(c) (d) sketch based interface using the gray-level co-occurrence matrix for
perceptual simplification of paper based scribbles in proceedings of the,”
in Eurographics Workshop on Sketch-Based Interfaces and Modelling,
Fig. 4. The results obtained after performing the proposed line tracking on S. Stahovich and Jorge, Eds., 2006, pp. 91 – 98.
a sample of scribble drawings [4] O. Hori, S. Tanigawa, and S. Shimotsuiji, “Precise Line Detection
form an Engineering Drawing Using a Figure Fitting Method based
on Contours and Skeletons,” in Proceedings of the 12th IAPR Int. Conf.
on Computer Vision and Image Processing, 1994, pp. 356 – 360.
The Kalman filter does not require the specification of a [5] W. Liu and D. Dori, “Sparse Pixel Vectorisation: An Algorithm and Its
Performance Evaluation,” IEEE Transactions of Pattern Analysis and
window size but performs the smoothing by considering the Machine Intelligence, vol. 21, no. 3, pp. 202–215, 1999.
measurement and process noise. The forward estimation step [6] X. Hilaire and K. Tombre, “Robust and Accurate Vectorization of
of the Kalman filter reduces the error by a mean value of 22%, Line Drawings,” IEEE Transactions on Pattern Analysis and Machine
Interpretation, vol. 28, no. 6, pp. 890–904, 2006.
while the backward smoothing step causes a further reduction [7] M. Röösli and G. Monagan, “Adding Geometric Constraints to the
of 21.3% such that the Kalman filtering reduces the error in Vectorisation of Line Drawings,” in Graphics Recognition - Methods
the medial points by a total of 43.3%. Hence, the Kalman filter and Applications, GREC’95, Lecture Notes in Computer Science, vol.
1072. Springer-Verlag, 1996, pp. 49–56.
gives better smoothing than the moving average filter without [8] K. Tombre, C. Ah-Soon, P. Dosch, G. Masini, and S. Tabbone, “Stable
requiring user defined parameters. and Robust Vectorization: How to Make the Right Choices,” Graphics
Fig. 4 shows that the proposed line location algorithm gives Recognition - Recent Advances, GREC’99, Lecture Notes in Computer
Science, Springer-Verlag, vol. 1941, pp. 3–18, 2000.
well placed lines that give a faithful representation of the [9] R. A. Katz and S. M. Pizer, “Untangling the Blum Medial Axis
scribble. The proposed vectorization represents the scribbled Transform,” International Journal of Computer Vision, vol. 55, no. 2-3,
drawing using few segments. This emerges since this line pp. 139–153, November 2004.
[10] L. Wenyin and D. Dori, “A Survey of Non-thinning Based Vectorization
tracking procedure updates the line parameters at each tracking Methods,” in Proceedings of the Joint IAPR International Workshops on
step, such that although the scribble is traced in a piecewise Advances in Pattern Recognition. Springer-Verlag, 1998, pp. 230–241.
linear manner, the tracking step and direction may adapt to [11] A. Sha’ashua and S. Ullman, “Structural Saliency: The Detection of
Globally Salient Structures using a Locally Connected Network,” in
the line curvature, enabling continuous tracking. Second International Conference on Computer Vision, Dec. 1988, pp.
321–327.
VI. C ONCLUSION [12] J. Zhou, W. Bischof, and T. Caelli, “Road Tracking in Aerial Images
Based on Human-Computer Interaction and Bayesian Filtering,” ISPRS
In this paper a sparse pixel based line location algorithm has Journal of Photogrammetry and Remote Sensing, 2006, pre-print ver-
been described. This line location algorithm differs from other sion.
sparse-pixel line tracking algorithms in that the sampling rate [13] P. Lallican and C. Viard-Gaudin, “A Kalman Approach for Stroke Order
Recovering from Off-line Handwriting,” in Proceedings of the Fourth
is determined adaptively, adjusting the tracking step to reflect International Conference on Document Analysis and Recognition, 1997,
the curvature of the drawing. Furthermore, Kalman filtering is pp. 519 – 522.
used in order to adjust the position of the data points, using [14] A. Lemaitre and J. Camillerapp, “Text Line Extraction in Handwritten
Document with Kalman Filter Applied on Low Resolution Images,”
the estimates of the scan line midpoints and the piecewise in Proceedings of the Second International Conference on Document
tracking to obtain a better estimate of the stroke’s medial Analysis for Libraries DIAL’06, 2006.
points. The proposed line tracking algorithm also introduces [15] P. S. Maybeck, Stochastic Models, Estimation and Control. Academic
Press Inc. (London) Ltd., 1979, vol. 1.
perceptual selection of the line tracking direction at junction [16] S. Sarkka, A. Vehtari, and J. Lampinen, “Time Series Prediction by
regions, allowing the tracking to follow the most salient path. Kalman Smoother with Cross-validated Noise Density,” Proceedings.
This, together with the constant adjustments in the tracking 2004 IEEE International Joint Conference on Neural Networks, vol. 2,
pp. 1653–1657, 2004.
direction, allow the proposed the line tracking algorithm to [17] W. Liu and D. Dori, “A protocol for Performance Evaluation of Line
trace the scribbled drawing using fewer line strokes, hence Detection Algorithms,” Machine Vision Applications, vol. 9, pp. 240 –
considerably reducing the computational times required in 250, 1997.
comparison to the other sparse-pixel tracking algorithms while
retaining the good quality line strokes obtained by these

View publication stats