Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V.

Strotov

AERIAL VEHICLES DETECTION AND RECOGNITION FOR UAV VISION SYSTEM

V.S. Muraviev1, S.A. Smirnov1, V.V. Strotov1
1
Ryazan State Radio Engineering University, Ryazan, Russia
Abstract
This article focuses on aerial vehicle detection and recognition by a wide field of view monoc-
ular vision system that can be installed on UAVs (unmanned aerial vehicles). The objects are
mostly observed on the background of clouds under regular daylight conditions. The main idea is
to create a multi-step approach based on a preliminary detection, regions of interest (ROI) selec-
tion, contour segmentation, object matching and localization. The described algorithm is able to
detect small targets, but unlike many other approaches is designed to work with large-scale objects
as well. The suggested algorithm is also intended to recognize and track the aerial vehicles of spe-
cific kind using a set of reference objects defined by their 3D models. For that purpose a computa-
tionally efficient contour descriptor for the models and the test objects is calculated. An experi-
mental research on real video sequences is performed. The video database contains different types
of aerial vehicles: airplanes, helicopters, and UAVs. The proposed approach shows good accuracy
in all case studies and can be implemented in onboard vision systems.
Keywords: aerial vehicles, object detection, contour descriptor, recognition
Citation: Muraviev VS, Smirnov SA, Strotov VV. Aerial vehicles detection and recognition for
UAV vision systems. Computer Optics 2017; 41(4): 545-551. DOI: 10.18287/2412-6179-2017-41-4-
545-551.
Acknowledgements: This publication has been prepared as a part of research carried out by
Ryazan State Radio Engineering University under the state contract 2.7064.2017/BCh.
Introduction Traditionally, some different approaches are used for
The unmanned aerial vehicles (UAVs) went through an recognition purposes, including hidden Markov models,
intensive development period in the last decade. In many feature points, tangent/turning functions, curvature maps,
applications they proved their reason for existence. Nowa- shock graphs, Fourier descriptors, etc. [4].
days the appearance of large numbers of UAVs raise new
problems such as autonomous navigation, early object detec-
tion and recognition, 3D scene reconstruction, collision
avoidance. It should be noted, that previously a lot of tasks
were solved by radar-based systems. They are reliable, but,
unfortunately, can’t be installed on the small aerial vehicles
because of high weight, size and energy consumption. So the
attention of researches is attracted by high resolution image
sensors. This paper is devoted to aerial vehicles detection
and classification by wide field of view monocular vision
system. This is essential for collision avoidance, autono-
mous drone swarm deployment, airspace patrol and monitor-
ing, in security applications.
A typical image of an aircraft at a range of several
kilometers is only a few pixels in diameter. A part of the
challenge is detecting such small targets in low signal to Fig. 1. Different types of observed aerial vehicles
background ratio. On the over hand, objects size con- They have their benefits and drawbacks, regarding
stantly grows as it approaches. Large object looks con- computational complexity, precision capabilities, imple-
trast but have too much detail that can lead the poor mentation issues, robustness and scalability. Other edge-
quality of object parameter estimation. The example of based approaches include Chamfer distance based meth-
observed types of aerial vehicles is shown in Fig. 1. ods for recognizing objects through smaller shape frag-
Well-known object detection algorithms are not always ments [7]. Complex algorithms based on machine learn-
invariant to the scale transform and are used primarily for ing [8] are developed actively, but they still have high
small target detection [2, 3]. However, some relatively re- computational costs.
cent research efforts look promising [4, 5]. The alternative Often aerial vehicles have homogenous brightness on the
approaches based on algorithm switching already applied for image and its shape information is more relevant. In this pa-
ground object detection, for example [6]. per, a relatively simple shape descriptor is used. It is compu-
However, the development of more reliable algo- tationally efficient and suited for onboard systems.
rithm for early object detection and confident recogni- This article focuses on aerial vehicles (airplanes, heli-
tion under different observation conditions is still an copters, UAVs) detection and recognition mostly in
important problem. cloudy background conditions. The main idea is to create

Computer Optics, 2017, Vol. 41(4) 545

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

a multi-step approach based on a preliminary detection, cause false detections. Next processing steps are per-
regions of interest (ROI) selection, object contour seg- formed to archive scale invariance, reject false detections,
mentation, contour descriptor calculation, object match- and refine object shape.
ing, and recognition. ROI selection. To archive invariance to scale trans-
In the next section object, detection algorithm is de- form gaussian image pyramid is created, and blob detec-
scribed in detail. Then contour descriptor evaluation, ob- tion algorithm described above is performed. Binary im-
ject matching and recognition are discussed. We present ages are formed at each scale of the pyramid. Filter masks
some experimental results for proposed approach ob- sizes are fixed, but coefficient k slightly increases with
tained on natural video sequences. image detail degradation. For each pyramid level binary
image is formed, and list of segments is created. Segment
1. Object detection
analysis at different scales is a part of the algorithm
Preliminary detection. The algorithm, that is used to which allows selection of regions of interest.
detect objects at preliminary step, should satisfy two re- The analysis starts from coarse image resolution and
quirements. It should be computational efficient and work goes to more detailed levels. Simple morphological oper-
well in cloudy and noisy environment. It can be assumed, ations are involved to reduce segment fragmentation on
that objects are more contrast than the underlying back- low resolutions. Bounding boxes for each segment are
ground. Neighboring pixels usually have similar bright- expanded on some value depending on initial size. Then
ness values and background have low spatial frequencies intersections between bounding boxes are searching at
in Fourier domain. In that case, objects with some as- different scales. Intersected regions must be counted and
sumptions can be describes as blobs. Spatial filters are excluded from the list. As a rule, large objects in the im-
typically used for blob detection to increase SNR and to age are more fragmented on detailed scale levels. This
get better results. property is used to specify large object location. Example
At first, the background must be estimated. Wherefore of binary mask of the test object on different levels of the
an observed image l(i, j) passes through the spatial filter with pyramid is shown in Fig. 2.
big size mask h2. Simultaneously l(i, j) is smoothed with
mask h1 of smaller size to average object brightness. To im-
prove performance, box filters are used. They use masks h1
and h2, which have dimensions (2q1+1)×(2q1+1) and
(2q2+1)×(2q2+1), q1 < q2 respectively:
h1 (m, n) = 1 / (2q1 + 1) 2 , m, n = −q1 , q1 ;
0, m, n = −q1 , q1 ; (1)
h2 (m, n) = 
( )
1 / (2q2 + 1) − (2q1 + 1) , otherwise.
2 2

After that the background estimation is subtracted

from filtered with mask h1 image:
d (i, j ) = l (i, j ) * h1 − l (i, j ) * h2 . (2)
Difference image d(i, j) contains objects and remain-
ing clutter with nearly gaussian zero mean spatial distri- Fig. 2. Image and binary mask at three levels of the pyramid.
bution. It can be concluded because of the large number Black rectangle corresponds to the ROI and is used for contour
of pixels in the image and application of a central limit segmentation
theorem. Taking into account the nature of the distribu- Thus, bounding boxes that are found on coarse resolu-
tion the thresholding procedure can be used to get object tion and have more intersections on higher resolutions are
binary mask b(i, j): probably related to objects and are treated as ROIs. The
size of all ROIs is transformed to one resolution scale.
1, d (i, j ) > k σ;
b(i, j ) =  (3) Remaining small segments found on original image are
0, otherwise, small targets. They can be described by its position, size
where k is a threshold coefficient and σ – standard devia- and average brightness. These characteristics are used for
tion of the difference image. matching based on minimization of relative differences of
However, application of the blob detection procedure object properties. Such small object can be tracked and its
in practice faces with a number of problems. Disad- velocity can help to increase recognition accuracy later.
vantages of the approach are explained by the locality of In opposite, large objects on binary image can look
spatial processing techniques. It is clear that the size of deformed. The segment centroid is often shifted, and that
the filter mask depends on the object size in the image. leads to significant errors in recognition. To achieve bet-
Large object is often fragmented, and it is imposable to ter results, object shape is restored at next step by pro-
correctly determine its shape. Besides, atmospheric turbu- cessing ROIs with contour segmentation algorithm robust
lence and background clutter such as contrast clouds to illumination changes.

546 Computer Optics, 2017, Vol. 41(4)

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

Contour segmentation. At this step, the more compli- flat. This effect is caused by δ(ϕ) due to ϕ(x, y) smooth-
cated segmentation procedure is performed in each ROI ing. The iterative search stops when the number of points
to estimate object contour. We choose active contour where level set function is close to zero ceases to vary no-
model as a powerful and flexible approach that can be ticeably [9].
used to precisely segment object boundary. More im- This method can deal with the detection of objects
portant that this approach grants that contour will be whose boundaries are dimmed or not necessarily defined
closed and won’t contain gaps. by gradient. It does not require image filtering and can ef-
The model is based on the Mumford-Shah functional ficiently process noisy images. Therefore, the true bounda-
minimization problem [9]. Let’s assume for simplicity ries are preserved and could be accurately detected. Addi-
that the images are continuous. The general form for the tionally, it can automatically detect interior contours with
Mumford–Shah energy functional for sensed image l(x, y) the choice of Dirac function approximation [10].
can be written as However, Chan-Vese model also has some draw-
backs: the unsuccessful segmentation of images with sig-
E MS (r , C ) = ∫∫ ( l ( x, y ) − r ( x, y) ) dx dy +
2
nificant intensity inhomogeneity, the sensitivity to the ini-
( x , y )∈ROI
(4) tial contour placement, and time-consuming iterative
+µ ∫∫ ∇r ( x, y ) dx dy + ν ⋅ length(C ),
2
solving procedure. In this work images are segmented on-
( x , y )∈ROI \ C ly in areas determined by ROIs, and are centered on ob-
where m and v are positive constants, l(x, y) – segmented jects in most cases. The influence of an image inhomoge-
image, С – object boundary curve. It becomes a difficult neity on segmentation results is noticeable for large-scale
problem to find С since r(x, y) is also an unknown func- objects but can be significantly reduced by image
tion in 2D coordinate space. Expression can be simplified downsampling in the gaussian pyramid. Thus, the main
if r(x, y) is a piecewise constant function that takes the drawbacks of the approach can be overcome.
value r1 inside С and r0 outside С. In that case energy Next subsection provides a description of object
functional (4) is reformulated as follows: recognition step of the algorithm.

E MS (r0 , r1 , C ) = λ1 ∫∫ ( l ( x, y ) − r0 ) ) dx dy + 2. Contour descriptor calculation

2

( x , y )∈outside ( C ) The contour descriptor is the number vector that is re-

(5)
+λ1 ∫∫ ( l ( x, y ) − r1 ) ) dx dy + ν ⋅ length(C ).
lates to the specified object contour. It is used for de-
2

( x , y )∈inside ( C )
creasing of the amount of information describing the ob-
ject contour. Also, the contours descriptor allows increas-
Expression (5) describes a Chan-Vese active contour ing the speed of the contour matching [11].
model [9, 10], where first term is the energy that corre- The proposed descriptor can be calculated using the
sponds to expansion force; the second is the energy that object binary image b or a contour C. In the first case af-
tends to compress the contour. The problem is to find the ter the image binarization we can extract external image
boundary of the object at which equilibrium is reached contour. Points of the contour are translated into polar
between two forces. The unknown curve C is replaced by coordinate frame with the frame center in the object cen-
the level set function ϕ(x, y), considering that ϕ(x, y) > 0 if troid. Obtained vector of polar coordinates is discretized
the point (x, y) is inside C, ϕ(x, y) < 0 if (x, y) is outside C, and subjected to the median filter.
and ϕ(x, y) = 0 if (x, y) is on C. Finally the minimization The result descriptor units can be calculated using the
problem is solved by taking the Euler–Lagrange equa- following equation:
tions and updating the level set function ϕ(x, y) by the
gradient descent method:    2πi π   
D(i ) = Fmed max  d ( P center , P  ,    , (8)

∂ϕ( x, y )    ND ND   
= δ(ϕ)( ν K (ϕ) − (l ( x, y ) − r0 ) 2 +
∂t (6)
where i = 1, N D – the number of the current descriptor
+(l ( x, y ) − r1 ) 2 ),
unit; ND – the total number of the descriptor units;
where r1 and r0 are average brightness values of object d(P1, P2) – Euclidian distance between P1 and P2; Pcenter –
and background respectively, δ(ϕ) – approximation of the position of the object centroid; P(α, ∆α) – any object
Dirac delta function, K(ϕ) – curvature of the curve. In or object contour point situated in sector of the circle that
transition from continuous (x, y) to discrete (i, j) coordi- is limited by the α±∆αangles (the circle is centered in
nate values equation (6) is transformed to Pcenter); Fmed{…} – the symbolic definition of the median
filtering operation.
ϕn +1 (i, j ) = ϕn (i, j ) + As the object contour is a close curve, it generates the se-
( )
(7)
+δ(ϕ) ν K n (ϕ) − (l (i, j ) − r0 (n)) 2 + (l (i, j ) − r1 (n)) 2 . ries of the descriptors that are shifted relative one to another
depending of the starting angle. The descriptor with the
At each n-th iteration, ϕ(x, y) is reinitialized to be the maximal D(1) unit is used as an object descriptor.
signed distance function to its zero level set. This proce- Steps of calculating contour descriptors are illustrated
dure prevents the level set function from becoming too in the Fig. 3.

Computer Optics, 2017, Vol. 41(4) 547

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

3. Object matching
Small targets are matched by minimizing relative dif-
ferences in average brightness, size position for object
candidates found in new frame. Contour coordinates are
very valuable for tracking and recognition purposes for
larger objects. However, information about contour coor-
dinates is excessive and values themselves are not invari-
ant to geometrical transformations. Therefore more rele-
vant contour descriptors are used.
Object matching is performed by minimizing the cri-
terion function:
 ND 
( ) ,
a)
Fcrit ( j ) = min  ∑ Dob (i − m ) − D j (i )
m∈M
 i =1  (9)
{ }
M = m : m = 1, N D , j = 1, N ,

where Dob – object contour descriptor found in the

previous frame, D j , j = 0, N is the descriptor of j object
candidate, N – number of objects in the current frame, m
– is the value of circular shift of a descriptor. Image rota-
tion results in circular shifts of the contour descriptor and
is taken into consideration in (9). Thus, matching process
is invariant to object rotation, scale and shift. Minimum
b) of Fcrit(j) for all j determines the most similar object.
4. Object recognition
The proposed object recognition algorithm consists of
two stages. The first stage is reference object database
preparation or learning. At this stage the reference obect
descriptors are calculated using 3D object models. The ref-
erence database includes a set of descriptors calculated for
a number of different object poses with different Euler an-
gles combinations. We suggest using the geosphere princi-
ple to distribute object poses on the sphere uniformly.
Since then stage includes a lot of complicated operation (as
3D model rendering), it produced preliminary [12].
The second stage of the algorithm is object recogni-
c) tion. It also is based on description of extracted image
contour and similar to the object matching algorithm.
This stage is performed in real time on the board.
The most probable pose is estimated as a result of
matching the contour descriptor of query image with
training descriptors. Descriptor matching is performed by
calculating the criterion function:
 ND 2
( )
f crit ( j ) = min  ∑ D0 (i − s ) − D j (i )  ,
s∈S
 i =1  (10)
{
S = s : s = 1, N D ,}
where D0 is the query image descriptor, Djis the de-
scriptor of current training image, and s is the value of
d) circular shift of descriptor.
Fig. 3. Steps of contour descriptor calculation: a – the input This criterion function provides rotation invariance of
query image of aircraft, b – binary image with the extracted the descriptor. Index of training descriptor corresponds to
contour and the example of the sector matched to the first a geosphere point. Therefore it determines Euler angles α
descriptor unit, c – binary image translated in polar coordinate and β. Let s0 to be the shift value that gives the minimum
frame, d – external contour descriptor value to the expression in square brackets in (11):

548 Computer Optics, 2017, Vol. 41(4)

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

 ND 2
( )
s0 ( j ) = arg min  ∑ D0 (i − s ) − D j (i )  ,
s∈S  i =1  (11)
{
S = s : s = 1, N D , }
Hence the value of angle γ is calculated by the formula:
2πs0
γ= . (12) a) b)
ND
As a result of calculating criterion function (11) for
every training descriptor we get vector of values of crite-
rion function

( )
M = f crit ( j ) | j = 1, N g . (13)

The measure of the similarity between captured object

and the k-th reference object is value Rk. It can be defined c) d)
as the minimal distance between sets of the object de-
scriptors as:
Rk = min ( f crit ( j ) ) , j = 1, N g . (14)
j

The recognition is processed by finding the least val-

ue of the:
match = arg min( Rk ), k = 1, K , (15)
k

where k – index of the most similar reference object, K –

the cardinality of the reference object set.
As the suggested algorithm is based on image contour
description, we meet the ambiguity problem. It happens e)
when calculated descriptor corresponds to more than one
orientation. For instance, topside and underside views of
an airplane will provide equal object contours and hence
descriptors. This problem must be taken in the account in
case of solving orientation estimation task, but it is not
important in case of recognition task solving.
Another algorithm problem is related to the defined
types of the object because of some image peculiarities. The
images of the helicopters often do not include the propeller.
It happens, for example, then the distance if far and the light
source is situated behind the observer. In this case the differ-
ence between the object contour descriptor and the reference
descriptors is inaccessibly high (Fig. 4e - f).
We propose to use two different models for the reference f)
descriptor generation. The first model of the helicopter in- Fig. 4. – Steps of contour descriptor calculation:
cludes the propeller, the others does not. The example imag- a – the render image of the helicopter with propeller
es of the aircraft obtained using this approach and the rela- b – the render image of the helicopter without propeller
tive contour descriptors are presented on Fig. 4. c – the binary image of the helicopter with propeller ,
d – the binary image of the helicopter without propeller,
5. Hardware implementation e – the contour descriptor of the helicopter with propeller,
Since the developed algorithm will be used in on- f – the contour descriptor of the helicopter without propeller
board vision systems, it must fit the system structure. The The object detection algorithm was designed for the
target systems consist of DSP or CPU as a control unit described system structure. The object recognition algo-
and a number of FPGAs as computing units. FPGAs are rithm cannot be performed onboard completely. The
used for performing the most of “heavy” operations such learning stage of this algorithm should be performed of
as spatial and temporal image filtering, geometric and external PC. It includes such specific operations as the
spectral transformations, template matching and thresh- 3D object rendering. Also, the learning stage has no time
olding, binary image marking. The DSP/CPU is used for restrictions. In contrast, the second stage of proposed al-
performing unique operations with small amount of data, gorithm is performed onboard in real time. The FPGA
FPGA dispatching and internal control. based vision system shows the most performance in case

Computer Optics, 2017, Vol. 41(4) 549

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

of pipelined processing. Therefore, we suggest that the In the work [16] the authors propose a mixed approach.
algorithms are suitable for Xilinx Virtex 5 of higher They use three types of indicators and a neural network.
FPGA based vision systems [12]. The result true positive ratio is between 82 % and 94 %. In
the work [17] the Markov random field based classificato-
6. Experimental research
ry is used. The result true positive ratio is between 88 %
The first goal of research is to determine the ability of and 95 %. In the work [18] the authors propose recognition
algorithm to localize objects at the distance of several of the military airplanes using wavelet-based descriptors.
kilometers. Video database contained 12 grayscale video The result true positive ratio is about 96 %.
sequences with 7 different types of aircraft, three types of The experiments were carried out on the same natural
UAVs and two helicopters. Object observed on cloudy image sequences that were used for the object detection
environment and in clear sky conditions. These sequences algorithm examination. The minimal aerial object area
were obtained from single TV or IR camera with a wide was 500 pixels. The maximal aerial object area was less
field of view. The size of objects varied from about 3×3 than 15 % of the image area.
pixels to 200×200 and even higher. Confident detection The reference object base includes 17 objects. The ob-
of objects in observed images affects the quality of algo- jects were defined by the 3D models. The sets of the ref-
rithm. The true positive ratio Pt and false negative ratio erence images were rendered for every model. The factor
Pf are measured for fixed detection algorithm parameters. 3 geosphere point distribution was used (92 points). The
Reference object position and size are determined in each light source was situated in front of the object. The ex-
video frame by visual inspection. Additionally the stand- amples of the object recognition are presented on Fig. 5.
ard deviation of object coordinate σc and size σs meas-
urement error are estimated.
To get more relevant results σs is divided on reference
size and expressed in percent. The results are summarized
in Table 1. The algorithm is less reliable in detecting hel-
icopters because of rotary wings that are not always dis-
tinguishable. In some cases the shape of the object varies
very rapidly due to the changes of the angle of view, a) b)
which also causes object misses.
Table 1. Object detection results grouped by type
Type of
aerial vehicles
Pd Pf σs, pixel σs, %

Airplane 0.95 0.06 3.6 5.4

Helicopter 0.89 0.12 6.3 10.5
UAV 0.93 0.08 2.4 3.2 c)
There are a lot of algorithms developed for aerial ob- Fig. 5. The example of the object recognition a – the object
source image, b – the object binary image with external
ject detection. In [13] authors adapted Viola-Jones algo- contour, c – the most similar reference object image
rithm for aircraft detection in video sequences. The prob-
ability of true detection ranged from 84.3 % to 89.1 %, The quality of the object recognition was estimated
depending on background conditions. The approach de- using the true positive recognition ratio metrics. This val-
veloped at Carnegie Mellon University [14] is focused on ue was averaged on the entry test video set. The results of
detection of small size unmanned aerial vehicles at a dis- the experiments are shown in the Table 2.
tance of about 3 miles from the image sensor. The algo- Table 2. Object recognition results
rithm provides detection probability of more than 90 % True
Estimated
with a false negative ratio not exceeding 5 %, but the al- № Object positive
object class
gorithm is not developed for large-sized object detection. ratio, %
A closer analogue of the developed algorithm is a multi- 1 Airbus 380 «Airplane» 96,1
step approach described in [15], which provides detection 2 LockheedC-130 «Airplane» 92,9
and classification of aerial objects. The algorithm demon- 3 Mi-172 «Helicopter» 98,3
4 MQ9 «UAV» 86,9
strated high detection quality; however, test video se-
5 MQ1 «UAV» 80,8
quences contained only aircraft on relatively low contrast 6 Mi-8 «Helicopter» 86,2
background. The effectiveness of the developed algo- 7 IL 76 (IR) «Airplane» 100
rithm is comparable with analogs, and in some cases it is 8 Lockheed C-5 «Airplane» 100
possible to achieve better results. 9 Cessna 172 «Airplane» 95,6
The second goal of the experimental research was to 10 Sukhoy SJ 100 «Airplane» 100
study the performance and accuracy of the proposed im- 11 Airbus 380 «Airplane» 87,2
age recognition algorithm in comparison with the 12 MQ9 «UAV» 89,2
known works. Average 92,8

550 Computer Optics, 2017, Vol. 41(4)

Aerial vehicles detection and recognition for UAV vision systems V.S. Muraviev, S.A. Smirnov, V.V. Strotov

Conclusion [8] Shen W, et al. Deepcontour: A deep convolutional feature

learned by positive-sharing loss for contour detection.
The proposed algorithms are suited for object detection
IEEE Conference on Computer Vision and Pattern Recog-
and recognition of aerial vehicles observed on cloudy back- nition (CVPR) 2015; 3982-3991. DOI:
ground under regular daylight conditions. Experiments show 10.1109/CVPR.2015.7299024.
that objects can be detected with good quality at the distance [9] Chan TF, Vese LA. Active contours without edges. IEEE
of several kilometers. Accuracy of matching and recognition Transactions on Image Processing 2001; 10(2): 266-277.
upon the average exceeds 90 % but depends on object type DOI: 10.1109/83.902291.
and orientation in space. The proposed algorithm is focused [10] Wang XF, Huang DS, Xu H. An efficient local ChanVese
on computational complexity reduction, and can be used in model for image segmentation. Pattern Recognition 2010;
airborne vision system installed on UAV. In future addition- 43(3): 603-618. DOI: 10.1016/j.patcog.2009.08.002.
[11] Arkin EM, Chew LP, Huttenlocher DP, Kedem K, Mitch-
al research work will be carried out to implement the algo-
ell JS. An efficiently computable metric for comparing po-
rithm in actual vision systems. lygonal shapes IEEE Transactions on Pattern Analysis and
References Machine Intelligence 1991; 13(3): 209-216. DOI:
10.1109/34.75509.
[1] Tirri AE, Fasano G, Accardo D, Moccia A, De Leis E. Ad- [12] Alpatov BA, Babayan PV, Ershov MD, Strotov VV. The
vanced Sensing Issues for UAS Collision Avoidance // implementation of contour-based object orientation estima-
Proceedings of the 2nd International Conference on Appli- tion algorithm in FPGA-based on-board vision system.
cation and Theory of Automation in Command and Con- Proc SPIE 2016; 10007: 100070A. DOI:
trol Systems 2012; 12-19. 10.1117/12.2241091.
[2] Lai JS, Mejias L, Ford JJ. Airborne vision‐based colli- [13] Petridis S, Geyer C, Singh S. Learning to detect aircraft at
sion‐detection system. Journal of Field Robotics 2010; low resolutions. Proceedings of the 6th International Con-
28(2): 137-157. DOI: 10.1002/rob.20359. ference Computer Vision Systems (ICVS 2008); 474-483.
[3] Nussberger A, Grabner H, Gool LV. Aerial object tracking [14] Dey D, Geyer CM, Singh S, Digioia M. A cascaded meth-
from an airborne platform. International Conference on od to detect aircraft in video imagery. The International
Unmanned Aircraft Systems (ICUAS) 2014; 1284-1293. Journal of Robotics Research 2011; 30(12): 1527-1540.
[4] Kovács L, Benedek C. Visual real-time detection, recogni- [15] Kovacs L, Benedek C. Visual real-time detection, recogni-
tion and tracking of ground and airborne targets. Proc SPIE tion and tracking of ground and airborne targets. Proc SPIE
2011: 7873: 787311. DOI: 10.1117/12.872314. 2011; 7873: 787311. DOI: 10.1117/12.872314.
[5] Kovács L, Kovács A, Utasi Á, Szirányi T. Flying target detec- [16] Rong HJ, Jia YX, Zhao GS. Aircraft recognition using
tion and recognition by feature fusion. Optical Engineering modular extreme learning machine. Neurocomputing 2014;
2012; 51(11): 117002. DOI: 10.1117/1.OE.51.11.117002. 128: 166-174. DOI: 10.1016/j.neucom.2012.12.064.
[6] Alpatov B, Korepanov S, Strotov V. A composite algo- [17] Li X-D, Pan J-D, Dezert J. Automatic aircraft recognition
rithm for variable size object tracking for high performance using DSmT and HMM. 17th International Conference on
FPGA-based on-board vision systems. Proc SPIE 2014; Information Fusion 2014.
9247: 92470A. DOI: 10.1117/12.2064058. [18] Karine A, Toumi A, Khenchaf A, El Hassouni M. Aircraft
[7] Shotton J, Blake A, Cipolla R. Multi-scale categorical ob- recognition using a statistical model and sparse representa-
ject recognition using contour fragments. IEEE Transac- tion //Proceedings of the International Conference on Big
tions on Pattern Analysis and Machine Intelligence 2008; Data and Advanced Wireless Technologies 2016; 49. DOI:
30(7): 1270-1281. DOI: 10.1109/TPAMI.2007.70772. 10.1145/3010089.3010134.

Authors’ information
Vadim Sergeevich Muraviev (b. 1981) received engineer qualification in Automation and Control Systems, candi-
date of Technical Science degree from Ryazan State Radio Engineering University (RSREU) in 2003 and 2010, respec-
tively. Currently he works as the associated professor of Automation and Information Technologies in Control depart-
ment (AITC) of RSREU. His research interests are image processing, pattern recognition, methods of classification and
optimization. E-mail: [email protected] .

Sergey Aleksandrovich Smirnov (b. 1986) received engineer qualification in Automation and Control Systems,
candidate of Technical Science degree from RSREU in 2008 and 2015, respectively. Currently he works as a leading
researcher in the AITC department of RSREU. His research interests are currently focused on computer optics, image
processing, pattern recognition, and control systems. E-mail: [email protected] .
Valery Viktorovich Strotov (b. 1980) received B.S., engineer qualification in automation and control systems,
candidate of technical science degree from RSREU in 2001, 2002 and 2009, respectively. He is a researcher (from
2002) and an associated professor (from 2009) of Automation and Information Technologies in Control department of
RSREU. His technical interests include image processing (image registration and stabilization, object detection, track-
ing and recognition) in industrial and on-board vision systems. E-mail: [email protected] .

Received May 24, 2017. The final version – August 14, 2017.

Computer Optics, 2017, Vol. 41(4) 551