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

net/publication/267155310

Automatic Scoring of Shooting Targets with Tournament Precision

Chapter · January 2012

DOI: 10.3233/978-1-61499-105-2-324

CITATIONS READS

0 1,477

2 authors, including:

Marcin Luckner
Warsaw University of Technology
47 PUBLICATIONS   87 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

LOKKOM View project

Variety, Veracity, VaLue: Handling the Multiplicity of Urban Sensors View project

All content following this page was uploaded by Marcin Luckner on 21 October 2014.

The user has requested enhancement of the downloaded file.

 Automatic Scoring of Shooting Targets with
Tournament Precision

Jacek Rudzinski and Marcin Luckner

Warsaw University of Technology, Faculty of Mathematics and Information Science,

pl. Politechniki 1, 00–661 Warsaw, Poland
rudzinskij@student.mini.pw.edu.pl
mluckner@mini.pw.edu.pl
http://www.mini.pw.edu.pl/~ lucknerm/en/

Abstract. This paper describes a computer vision based automatic

scoring system of shooting targets. The system estimates scoring with
a professional tournament precision, but is dedicated to amateur shoot-
ers and can work with photos taken by amateur cameras and mobile
devices. The automatic scoring issue is divided into three problems: a
target detection, a holes detection, and a hole analysis. The target is
detected on the base of a bull–eye localization. The holes detection bases
on the Hough transformation. The holes analysis localizes a position
of hole’s center. The position relative to detected scoring sections is a
base for scoring. The proposed algorithm detects holes with 99 percent
accuracy. An elimination of false positives results reduces the level of
accepted holes to 92 percents. The average error for the automatic score
estimation is 0.05 points. The estimation error for over 91 percent holes
is lesser than a tournament–scoring threshold.

Keywords: Computer vision, Hough transform, Pattern recognition,

Score estimation

1 Introduction

Automatic scoring of shooting target is done by an expensive equipment such as

shooting targets scanners [6], optical evaluation systems [2], electronic shooting
targets [12], or acoustic systems [3].
As a cheaper alternative a computer vision based automatic scoring system
can be used. However, existing systems also need a special equipment as high–
resolution digital camera [1] or DH–CG320 capture card [8]. Moreover, there are
projected as stationary systems for professional shooters.
Our aim was a creation of a professional system that can be used by amateur
shooters. The professional system should estimated a score on the level accepted
by an International Shooting Sport Federation (ISSF) as it is shown in Fig. 1.
Results obtained by the system should be at least as well as given by computer
vision stationary systems but a cost of the system should be within reach for
amateur shooters.
2 Automatic Scoring of Shooting Targets with Tournament Precision

Fig. 1. A professional scoring with a tournament precision. The score is 8.8 in this
case.

Amateurs cannot afford high resolution digital systems. For that reason, the
system was projected for the following minimal requirements:
– an image of a shooting target should be greater than 0.5 Mpix,
– at least whole bull–eye should be visible,
– an angle between a normal of a target plane and an optical axis should be
no more then 30 degree,
– a less should have a small spherical aberration,
– a focus should be comparable with a matrix diagonal,
– a target should be flat, cannot be folded,
– a target should be made from a good quality paper and meets ISSF stan-
dards,
– a pellet should be used for shooting,
– annotations on a target should not be similar to holes.
Described requirements can be satisfied by any amateur camera and a most
of mobile devices.
The described system has three components. The first one is a target detec-
tion, which is described in section 2. The next step is detection of holes and this
problem is discussed in section 3. The last step is analysis of holes. In this step
scoring points of holes are localized. Details are given in section 4.
The algorithm was tested on several amateur photos of targets and section 5
gives the experimental results. The paper is concluded in section 6.

2 Target detection
A target detection includes a localization of its components. The main element
of the target is a bull–eye, which is a central located black circle on a white
background. Rings are additional white (inside the bull–eye) and black (outside
the bull–eye) circles, which separate scoring sections.
 Automatic Scoring of Shooting Targets with Tournament Precision 3

Fig. 2. Amateur photos of targets. A photo on the right has an inconsistent illumina-
tion.

2.1 Bull–eye detection

A target has a black circle bull–eye on a white background. This pattern should
be easy to detect, for example by a binarization. However, amateur photos (such
as given in Fig. 2) are not taken under laboratory conditions. The photos can
be taken at home, for example to archive a daily shooting results. Such factors
as an inconsistent illumination can make the binarization based detection hard.
Such example, when a global threshold for the binarization cannot be found, is
given in Fig. 3. For that reason, an edges detection based algorithm is proposed.

Fig. 3. Photos converted to binary images. Bull-eyes in a photo on the right are blurred.

The algorithm works on pictures in reduced resolution. For each pixel p a

maximum brightness difference is calculated between the pixel and its eight–
point neighborhood P = {p1 . . . p8 } as

γp = max |lum(p) − lum(px )|, (1)

px ∈P
4 Automatic Scoring of Shooting Targets with Tournament Precision

Fig. 4. An edges detection based approach. Detected targets are marked with squares

where lum(p) is a p point brightness.

Next, a condition
lum(p) − γp > 0, (2)

is checked for each pixel that lies on an edge of detected shape. If almost all
pixels satisfy the condition then the shape is a potential bull–eye. The second
condition that should be satisfied is

max(W, H)
< 2, (3)
min(W, H)
where W and H are width and hight respectively. The results of the algorithm
for the amateur photos (Fig. 2) are shown in Fig. 4.
The detected bull–eye brings local information about a threshold for a bi-
narization. After the binarization a set of point on the bull–eye edge will be a
source of information about a geometry of rings projections.

2.2 Rings detection

The set of point on the bull–eye edge determines an approximated shape of the
main ring. After corrections, selected points can be used in a parametrization of
the ring projection.
The bull–eye was detected in the low resolution image. A ring geometry
should be detected in the original photo. For that reason a detected points are
only an approximation. It can be assume that a point from the ring lies on the
line determined by an approximated point and the center of the target. Moreover,
a distance between the points should not be more than 16 of a distance between
rings.
These conditions define a segment that includes the ring point. For the seg-
ment, the value V from HSV color model is calculated in each point. The segment
has the minimum, the maximum, and the average values labeled as Vmin , Vmax ,
 Automatic Scoring of Shooting Targets with Tournament Precision 5

(a) (b) (c) (d)

Fig. 5. Holes detection. 5(a) A result of the Prewitt operator. 5(b) Hough transforma-
tion results with a significant number of false positives. 5(c) A result of the Prewitt
operator with erased rings. 5(d) Hough transformation results with a reduced number
of false positives.

and Vavg respectively. For a given threshold α ∈ [0, 1] a point from the ring pring
is calculated as

pmin if κ > 1 − α

pring = pmax if κ < α (4)

pdif f if α < κ > 1 − α,


where
Vavg − Vmin
κ= . (5)
Vmax − Vmin
The selected point can be the most brightness pmax , the darkest pmin , or with
the highest difference between neighbors pdif f .
Some points from the set can be taken as a part of the ring by mistake.
To avoid such situation, several ellipses is generated from a random subsets of
points [9]. The created ellipses are evaluated with distance criteria. The best
subset is used in the future works.
A least squares method can be used for points based ellipse parametriza-
tion [10]. However, this method is numerical unstable and a stable solution was
used instead [11].
All rings are projections of circles from a target. Radius of the circles are
known. When parameters of the projected bull–eye are calculated then param-
eters of all ellipses can be also obtained.

3 Holes detection
There should be about ten holes in the target. A hole can be estimated by an
ellipse. Ellipses can be detected using the Hough transformation [7].
The Hough transformation based ellipse center detection is a voting proce-
dure. The point that gets the biggest number of votes wins. All results of Hough
transformations can be ordered by a number of votes given to the center. Among
them twenty best results is taken for the future analysis.
6 Automatic Scoring of Shooting Targets with Tournament Precision

Fig. 6. Holes analysis. From the top: shapes created by the flood fill algorithm, detected
edges, results of Hough transformation, and estimated centers of holes. The most right
example is a false positive.

As an input for the Houg transformation results of the Prewitt operator can
be used. The Prewitt operator is an edges detector that gave the best results
among tested. However, rings, numbers, and other marks in the target have a
significant influence on Hough transformation results. In Fig. 5 an input image
5(a) and Hough transformation results 5(b) are shown. Among the best results is
a significant number of false positives. Moreover, several existing holes is ignored
because their results are placed outside the first twenty.
The results can be improved when detected rings (section 2.2) are erased
from the image. In Fig. 5 an input image with erased rings 5(c) and Hough
transformation results 5(d) are shown. This time a number of false positives is
scanty and all holes are detected.

4 Holes analysis

The main aim of a holes analysis is a precision localization of holes centers.

Moreover, it can be used to reduce a number of false positive results.
The analysis of detected holes has the following steps. First, a shape of hole is
determined by a flood fill algorithm. If the algorithm is used on images different
than binary then a color tolerance has to be defined. The distance between
colors can be calculated as a maximum difference between components of the
RGB color model.
In the next step, the Prewitt operator is used to detect edges. The edges
are an input for the Hough transformation that can be used to detect an ellipse
center [4,5]. All steps are shown in Fig. 6.
Similarly as in the section 3 all results of Hough transformations can be
ordered by a number of votes given to the center. A rejection threshold is calcu-
 Automatic Scoring of Shooting Targets with Tournament Precision 7

(a) (b)

Fig. 7. The shooting targets. 7(a) The target with one missed hole. 7(b) The target
with detected all holes and an angle between a normal an optical equals 30 degree.

lated as 60 percent of the best result. All results under the threshold are rejected.
Remaining holes are classified as valid.

5 Results
In this section tests results are described. Two aspects were evaluated. Number
of detected holes and a quality of estimated score.

5.1 Holes detection

The tests were done for 14 shooting targets with 152 holes. The examples of the
shooting targets are given in Fig. 7. The hole detection algorithm marked 208
objects. Among this number 150 were correctly detected holes and 58 were false
positives. Details are given in Table 5.1. For each target a number of holes (ALL)
is shown as well as a number of detected ones (T P ). The quality of detection is
TP
calculated as ALL . On the other hand, a number of false positive (F P ) decision
is shown and its signification calculated as F PF+T
P
P.
The detection results are very good. About 99 percent of holes were detected.
The results are better than published in [1]. However, among all positive decisions
28 percent were false positive. For that reason the future analysis is needed.
The results of detection were used as input data for an analysis. The analysis
rejected all false positive inputs except two. A negative aspect is that 13 correct
holes were also eliminated. Details are given in Table 2. For each target a number
of positive (P I) and negative (N I) inputs from the detection is given. The quality
of the analysis results is estimated by a classification error for each group of
inputs.
The analysis rejected 97 percent of false positives inputs. However, about 8
percent of holes is rejected at the same time.
The total quality of the algorithm calculated as a percent of detected holes
multiplied by a percent of accepted in the analysis is 92 percent.
8 Automatic Scoring of Shooting Targets with Tournament Precision

Table 1. Results of holes detection

Shooting HolesTrue T P False FP

target positive ALL positive (F P + T P )
1 9 9 1.00 7 0.44
2 10 10 1.00 5 0.33
3 10 10 1.00 6 0.38
4 9 9 1.00 1 0.10
5 13 13 1.00 2 0.13
6 11 11 1.00 4 0.27
7 13 13 1.00 2 0.13
8 13 13 1.00 2 0.13
9 12 11 0.92 3 0.21
10 13 13 1.00 2 0.13
11 9 9 1.00 7 0.44
12 9 9 1.00 6 0.40
13 11 10 0.91 6 0.38
14 10 10 1.00 5 0.33
Total 152 150 0.99 58 0.28

5.2 Score estimation

When a center of a hole in known then its score can be calculated. A pellet
diameter is constant and can be used to determine a radius of the hole. The
scoring point lies on a circumference determined both the hole’s center and the
pellet’s radius. The point that lies nearest the target center should be selected.
The score is given by a value of the smallest ring that includes the point.
This method of estimation was tested on 148 holes. The results were com-
pared with a human estimation. The distribution of errors is given in Fig. 8.
The average error is 0.05 points. For over 91 percent of holes, the error is less
than 0.1 points, which is a typical scoring precision.

5.3 Comparison with related methods

The results cannot be compared directly with the other related method projected
as stationary systems for professional shooters. The paper [1] is limited to a holes
detection and a score calculation is not presented. However, the accuracy of the
holes detection (98.3 percent) is similar to our results (99 percent) before the
rejection. The results after rejection cannot be compared because a problem of
false positives is not discussed in the paper.
The second paper [8] presents only a total score, which is obtained as a result
of an artificial target scoring and an automatic target scoring. The difference
between them is 0.1 points and this level is presented by the authors as an
obtained error. Errors for separated holes are not given. The scoring precision
in our solution is similar. The average error is 0.05 points and the error is less
than 0.1 for over 91 percent of holes.
 Automatic Scoring of Shooting Targets with Tournament Precision 9

Table 2. Results of holes analysis

Shooting Input Positive Negative False F P False F N

target input input positive N I negative P I
1 16 9 7 1 0.14 1 0.11
2 15 10 5 0 0.00 1 0.10
3 16 10 6 0 0.00 1 0.10
4 10 9 1 0 0.00 0 0.00
5 15 13 2 0 0.00 1 0.08
6 15 11 4 0 0.00 1 0.09
7 15 13 2 0 0.00 1 0.08
8 15 13 2 0 0.00 0 0.00
9 14 11 3 0 0.00 0 0.00
10 15 13 2 0 0.00 1 0.08
11 16 9 7 0 0.00 1 0.11
12 15 9 6 0 0.00 0 0.00
13 16 10 6 0 0.00 1 0.10
14 15 10 5 1 0.20 1 0.10
Total 208 158 58 2 0.03 13 0.08

In sum, our solution dedicated to amateur shooter gives similar results as

stationary systems for professional shooters, but not all aspects of the scoring
can be compared.

6 Conclusions

In the paper an issue of automatic scoring of shooting target was described. The
proposed solution divides the issue into three problems: a target detection, a
holes detection, and a hole analysis.
The target detection based on a set of several basic algorithms works de-
pendable. The detection localizes targets on amateur photos if only minimal
requirements are satisfied.
The implemented the Hough transformation based hole detection finds 99
percent of holes in photos of a resolution at least 1MPix. Among detected holes
are false positive responses. Additional analysis eliminates most of them, but the
detection of positive inputs is reduced to 92 percent. An improvement of this
analysis is an aim of the future works.
The tests proved that the average error for the automatic score estimation
is 0.05 points. For over 91 percent of holes the error is less then 0.1 points,
which is a typical scoring precision and similar to results of stationary systems
for professional shooters.
Low minimum requirements for processed images create a possibility of de-
veloping the described application on mobile devices. In such case, it can be a
handy application for amateur shooters.
 10 Automatic Scoring of Shooting Targets with Tournament Precision

Fig. 8. Estimation errors distribution

References
1. Ali, F., Bin Mansoor, A.: Computer vision based automatic scoring of shooting
targets. In: Multitopic Conference, 2008. INMIC 2008. IEEE International. pp.
515 –519 (dec 2008)
2. Aviatronic Ltd.: SmartSCORE Users Manual (2005),
http://www.smartscore.hu/kep/users_manual_en.pdf
3. C. Sanctuary, A. Sean, S.R.H.: Remote strafe scoring system. United States Patent
4813877 (1989)
4. Chia, A., Leung, M., Eng, H.L., Rahardja, S.: Ellipse detection with hough trans-
form in one dimensional parametric space. In: Image Processing, 2007. ICIP 2007.
IEEE International Conference on. vol. 5, pp. V –333 –V –336 (16 2007-oct 19
2007)
5. Davies, E., Barker, S.: An analysis of hole detection schemes. pp. 285–290 (1990)
6. DISAG-INTERNATIONAL: DISAG RM IV operating instruction (2005),
http://www.disag.de/download/manuals/rmiv_en.pdf
7. Duda, R.O., Hart, P.E.: Use of the hough transformation to detect
lines and curves in pictures. Commun. ACM 15(1), 11–15 (Jan 1972),
http://doi.acm.org/10.1145/361237.361242
8. Fan, X., Cheng, Q., Ding, P., Zhang, X.: Design of automatic target-scoring system
of shooting game based on computer vision. In: Automation and Logistics, 2009.
ICAL ’09. IEEE International Conference on. pp. 825 –830 (aug 2009)
9. Fichler, M., Bolles, R.: Random sample consensus: A paradigm for model fitting
with applications to image analysis and automated cartography (1980)
10. Fitzgibbon, A. W.and Pilu, M., Fisher, R.B.: Direct least-squares fitting of ellipses
21(5), 476–480 (May 1999)
11. Halir, R., Flusser, J.: Numerically Stable Direct Least Squares Fitting of Ellipses
(1998), http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.1.7559
12. SIUS ASCOR: Electronic Scoring systems (2010),
http://www.sius.com/downloads/docu/Usermanual_System7_e.pdf

View publication stats