Generic Fourier Descriptors for Autonomous UAV Detection

Eren Unlu

, Emmanuel Zenou

and Nicolas Riviere

ISAE-SUPAERO, 10 Av. Edouard Belin, Toulouse, France

ONERA, 2 Av. Edouard Belin, Toulouse, France

Keywords:

Aerial Surveillance, Drone Detection, Generic Fourier Descriptor, Shape Descriptors, Object Recognition.

Abstract:

With increasing number of Unmanned Aerial Vehicles (UAVs) -also known as drones- in our lives, safety and

privacy concerns have arose. Especially, strategic locations such as governmental buildings, nuclear power

stations etc. are under direct threat of these publicly available and easily accessible gadgets. Various methods

are proposed as counter-measure, such as acoustics based detection, RF signal interception, micro-doppler

RADAR etc. Computer vision based approach for detecting these threats seems as a viable solution due

to various advantages. We envision an autonomous drone detection and tracking system for the protection

of strategic locations. In this work, 2-dimensional scale, rotation and translation invariant Generic Fourier

Descriptor (GFD) features (which are analyzed with a neural network) are used for classifying aerial targets

as a drone or bird. For the training of this system, a large dataset composed of birds and drones is gathered

from open sources. We have achieved up to 85.3% overall correct classiﬁcation rate.

1 INTRODUCTION

Recent advances in Unmanned Aerial Vehicles (UAV)

-usually preferred to refer publically as drones- indus-

try made these devices highly accessible to all kinds

of civilians. Licensing and regulating drone utiliza-

tion has lagged behind this rapid expansion of the in-

dustry. Besides their numerous advantages, drones

have a huge potential to be misconducted intention-

ally or non-intentionally. First of all, commercial

UAVs, even the cheapest and smallest ones can be

easily converted to a weapon for terrorism by attach-

ing explosives. Moreover, irresponsible utilization of

these devices may cause also panic and fatal acci-

dents. For example, during a football game in Texas,

a drone has entered the stadium’s airspace, causing

large panic for spectators and law enforcement units.

The responsible security forces rested helplessly as

they have lacked the proper surveillance and counter-

measuring equipment (Humphreys, 2015). In addi-

tion to this, commercial drones being ﬂown around

airports raising concerns for civilian aviation secu-

rity. Numerous incidents caused by drones were re-

ported by aviation ofﬁcers in recent years (Wild et al.,

2016).Also personal privacy issue and preventing in-

dustrial or governmental espionage is another serious

problem (Villasenor, 2013).

These incidents have caused a paradigm shift

for the governmental defence and public security

strategies, as these crafts cannot be detected efﬁ-

ciently with conventional methods, such as RADARs

etc. due to their size and small electromagnetic sig-

natures (Peacock and Johnstone, 2013). Industry

and academy have focused on new kind of counter-

measure methods, where a solid consensus still is

not apparent. The mostly used methods for small to

medium sized UAV detection are RF detection (de-

tecting the RF signals for control between the oper-

ator and the drone), acoustics (detecting certain spe-

ciﬁc sounds emitted from the rotors of the drone), X-

band RADAR, micro-doppler signature (RADAR for

detecting small moving objects like drones) and the

optical methods (detection by computer vision) (Yoon

et al., 2017)(Franklin and Hearing, 2016)(Solodov

et al., 2017). Each of these methods has its own ad-

vantages or drawbacks. For instance, acoustics based

detection with directional microphone arrays has a

relatively low range of approximately 250 meters.

And they are highly sensitive to background noise,

which is a complicated problem especially in urban

areas. RF signal interception disregards the fact that

certain drones may not be controlled via wireless con-

nection, but may have been preprogrammed to follow

a certain route. Micro-doppler and X-band radar strat-

egy often causes high number of false alarms, caused

by birds, background clutter etc. Thus, they are gen-

550

Unlu, E., Zenou, E. and Riviere, N.

Generic Fourier Descriptors for Autonomous UAV Detection.

DOI: 10.5220/0006680105500554

In Proceedings of the 7th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2018), pages 550-554

ISBN: 978-989-758-276-9

erally accompanied an additional detection and iden-

tiﬁcation apparatus.

Among these strategies, detection by computer vi-

sion distincts itself with its efﬁciency and robustness.

It can also be applied to infrared cameras, thus pro-

viding night time operation capability. In this paper,

we present a new method for drone detection with

computer vision. This method can also be applied to

infrared imagery. We assume that a moving object

detecting background subtraction algorithm is contin-

uously used and detected blobs’ binary silhouette is

examined with the Generic Fourier Descriptor (GFD)

based algorithm to detect drones. In second section,

we refer to existing drone detection technologies by

computer vision approach and in the third section we

explain the Generic Fourier Descriptor (GFD) pro-

posed by (Zhang and Lu, 2002) and our approach

to detect drones. One of the most important chal-

lenges for drone detection is avoiding false alarms

caused by the birds (G

okc¸e et al., 2015). Thus, we

have used ﬂying bird silhouettes also in our algo-

rithm for better discrimination capability. Next, in the

fourth section we explain our experimentation method

and results. Thanks to this approach, we could have

achieved 85.3% overall correct classiﬁcation rate be-

tween drone and bird silhouettes.

2 COMPUTER VISION FOR UAV

DETECTION

As mentioned previously, computer vision for de-

tecting drones is a more robust, feasible and effec-

tive method compared to other existing ones. Con-

volutional Neural Networks (CNNs) are the state-of-

the-art method for object detection and identiﬁcation,

which has not a long history (Ciresan et al., 2011).

It is a deep learning technique, which autonomously

learns the optimal features for classiﬁcation by im-

agery, thus does not depend on human crafted features

(Simard et al., 2003). Recently, for computer vision

based detection various authors have oriented them-

selves to CNNs. Among these, we see (Schumann

et al., 2017)(Saqib et al., 2017)(Aker and Kalkan,

2017), which are using very similar approach for

CNNs, however with different architectures. CNNs

may be the most recent and state-of-the-art solution in

the literature, however they require extensive compu-

tational cost, especially for training. In addition, their

accuracy may be still low for certain circumstances

such as low resolution, insufﬁcient dataset etc.

Rather than CNNs, (Unlu et al., 2017) uses SURF

based keypoint features of grayscale drone, bird and

background image patches. The authors propose a

new kind of extended bag-of-words (BoW) approach

for classiﬁcation. In this paperwork, we propose a

GFD based approach for classifying image patches

composed of birds and drones similar to those in

(Unlu et al., 2017).

3 GFD BASED DRONE

DETECTION

3.1 Generic Fourier Descriptors

Fourier Descriptors have been used as an efﬁcient

shape descriptor (Persoon and Fu, 1977). The dis-

tances of each contour pixel to the center of mass of

the 2D object silhouettes is represented as a vector.

Fourier Transform of this vector gives a unique de-

scription of the shape as the transform itself is shift,

scale and rotation invariant.

Generally, the lower frequencies of the transform

contains more information on the major structural

parts of the object. If we interpret the mechanism of

the algorithm, we can state that higher frequencies of

the transform correspond to the more intensive ripples

on the contour.

However, even this approach can differentiate

non-similar silhouettes with high efﬁciency, the clas-

siﬁcation performance degrades as the contours get

similar. In addition to this, as mentioned previously,

this algorithm only considers the shape of the out-

side contours. However, the form of the silhouettes

can contain very important and distinctive informa-

tion such as holes etc. An approach taking into ac-

count the complete silhouette shall be more robust to

noise which can miss certain number of pixels.

Generic Fourier Descriptor (GFD) is a method

proposed by (Zhang and Lu, 2002), which takes into

account the 2D object silhouette in contrast the uni-

dimensional Fourier Descriptors. The idea is to ﬁrst

raster and transform the pixels of the silhouette to po-

lar coordinates with chosen angular and radial reso-

lutions. Normalized 2D Fourier transform (Eq. 1) of

this rastered function generates two dimensional ma-

trix which we use as the representation of the shape.

When this result is being used for classiﬁcation with

various algorithms, it is vectorized (Zhang and Lu,

2002).

GFD(R, T ) =

∑

( f (r, θ))e

(− j2πr

)+2π

(1)

As it is normalized, this method is intrinsically

scale invariant. And due to polar mapping by tak-

ing the center of mass as the origin, it is also translate

Generic Fourier Descriptors for Autonomous UAV Detection

551

Angular

resolution

Radial

resolution

2D GFD transformation

Figure 1: The GFD transformation of a 2D object silhou-

ette.

and rotation invariant just like the regular Fourier De-

scriptors. Fig. 1 illustrates the GFD calculation of a

2D object silhouette.

3.2 Aerial Surveillence by Using GFD

The binary silhouettes of objects are composed of 1s

(where the pixel corresponds to object) and 0s (where

pixel corresponds to background) in an arbitrary size

frame. These are determined by the moving object

detector (background subtraction). Then, we classify

the object by using Generic Fourier Descriptor (GFD)

features and a neural network. In order to train our

system, we have created a dataset, composed of im-

ages of ﬂying birds and drones, which are acquired

from open sources.

To seperate the object pixels from the background,

a special image segmentation algorithm is applied.

We have composed the dataset from the images,

where the object is darker than the background (i.e.

sky). The images are chosen to be relatively low reso-

lution in order to reﬂect the target case, where the au-

tonomous tracker detect a small ﬂying object in wide

angle. All images are converted to gray scale and

rescaled to 64x64 pixels. Fig. 2 shows few of the

images from the dataset both for drones and birds.

Region Growing algorithm is chosen as the image

segmentation algorithm to seperate the object silhou-

ettes from the background in the images due to its efﬁ-

ciency (Adams and Bischof, 1994). Region Growing

is a method, where pixel neighborhoods are evaluated

in an iterative manner, starting from an initial seed

point. Over course of the algorithm, the pixels are

deﬁned as background or foreground, by applying a

clustering criterion.

After we have seperated the object silhouettes, we

have applied a 2D GFD algorithm with 16 radial and

9 angular resolutions (therefore, we have feature vec-

tors composed of 144 scalar values for each object.).

Table 1: Confusion matrix for the classiﬁcation of bird and

drone silhouettes in test set.

Bird

207

68.3%

9.6%

87.7%

12.3%

Output

Class

Drone

4.6%

17.5%

79.1%

20.9%

93.7%

6.3%

64.6%

35.4%

85.8%

14.2%

Target

Class

Bird Drone

In order not to lose any shape information, we have

not applied any morphological operations after im-

age segmentation phase. For most of the cases, after

image segmentation there is no need for further pro-

cessing to acquire the true binary silhouettes of the

objects. However, in case there are more than one

disconnected pixel groups, the algorithm chooses the

largest pixel group as the true binary silhouette of the

object. To apply GFD, the binary silhouettes of the

objects have to be centered in a 2D plane, where their

center of mass is the origin. Note that, as GFD is a

scale and rotation invariant transform, there is no need

for rescaling or rotation for the silhouettes. Fourier

Transformation results are normalized and reshifted

before further processing. Following this, we have

created a neural network composed of approximately

6000 neurons to classify GFD features in to birds and

drones.

4 EXPERIMENTATION AND

RESULTS

We have used 410 drone images and 930 bird images.

A 5-fold approach is followed, where 4/5 of the sam-

ples are always used for training and the 1/5 of the

samples is used for testing. In addition to this, to as-

sure the validity of the experiments, the training sam-

ples are again divided in an additional 5-fold manner,

where the test group is used for developing regulariza-

tion parameters during the optimization of the neural

network.

We have acquired an overall 85.3% accuracy on

the test groups. The Table 1 shows the confusion ma-

trix for a test group (0 : birds and 1 : drones). As

it can be seen, the GFD based algorithm is especially

effective at detecting bird shapes, with a true rate of

93.7%. However, we see that only 64.6% of drones

are correctly identiﬁed. The overall accuracy shows

that GFD is a novel 2D shape descriptor for discrim-

inating bird and drone silhouettes for an autonomous

drone surveillence system.

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

552

Birds

Drones

Background

Figure 2: A few examples of the 64x64 grayscale bird, drone and background patches from the images we have collected

from the internet.

5 CONCLUSION

We have showed that GFD can be used as a valuable

shape descriptor along with a complex neural network

for an autonomous drone surveillence system. To the

best of our knowledge, using GFD with neural net-

works for an aerial surveillence system has not been

visited. The results may be further augmented by us-

ing a larger dataset and additional features.

This approach may compete with the currently

preferred CNN based algorithms, which require ex-

tensive computational power and very large datasets,

while providing no information on the curvature of

the object. Another advantage of GFD based algo-

rithm is its potential ability to be used for motion

based changes. This can be performed by analyzing

temporal changes of GFD features. For CNNs this

analysis is cumbersome and needs other type of deep

learning techniques.

ACKNOWLEDGEMENTS

This work is supported by the French Ministry of De-

fence.

REFERENCES

Adams, R. and Bischof, L. (1994). Seeded region growing.

IEEE Transactions on pattern analysis and machine

intelligence, 16(6):641–647.

Aker, C. and Kalkan, S. (2017). Using deep networks for

drone detection. In Advanced Video and Signal Based

Surveillance (AVSS), 2017 14th IEEE International

Conference on, pages 1–6. IEEE.

Ciresan, D. C., Meier, U., Gambardella, L. M., and

Schmidhuber, J. (2011). Convolutional neural net-

work committees for handwritten character classiﬁca-

tion. In Document Analysis and Recognition (ICDAR),

2011 International Conference on, pages 1135–1139.

IEEE.

Franklin, J. and Hearing, B. (2016). Drone detection and

classiﬁcation with compensation for background clut-

ter sources. US Patent App. 15/360,069.

okc¸e, F.,

Uc¸oluk, G., S¸ ahin, E., and Kalkan, S. (2015).

Vision-based detection and distance estimation of mi-

cro unmanned aerial vehicles. Sensors, 15(9):23805–

23846.

Humphreys, T. (2015). Statement on the security threat

posed by unmanned aerial systems and possible coun-

termeasures. Oversight and Management Efﬁciency

Subcommittee, Homeland Security Committee, Wash-

ington, DC, US House.

Peacock, M. and Johnstone, M. N. (2013). Towards detec-

tion and control of civilian unmanned aerial vehicles.

Persoon, E. and Fu, K.-S. (1977). Shape discrimination us-

ing fourier descriptors. IEEE Transactions on systems,

man, and cybernetics, 7(3):170–179.

Saqib, M., Khan, S. D., Sharma, N., and Blumenstein, M.

(2017). A study on detecting drones using deep con-

volutional neural networks. In Advanced Video and

Signal Based Surveillance (AVSS), 2017 14th IEEE In-

ternational Conference on, pages 1–5. IEEE.

Schumann, A., Sommer, L., Klatte, J., Schuchert, T., and

Beyerer, J. (2017). Deep cross-domain ﬂying object

classiﬁcation for robust uav detection. In Advanced

Video and Signal Based Surveillance (AVSS), 2017

14th IEEE International Conference on, pages 1–6.

IEEE.

Simard, P. Y., Steinkraus, D., Platt, J. C., et al. (2003). Best

practices for convolutional neural networks applied to

visual document analysis. In ICDAR, volume 3, pages

958–962.

Solodov, A., Williams, A., Al Hanaei, S., and Goddard, B.

(2017). Analyzing the threat of unmanned aerial vehi-

Generic Fourier Descriptors for Autonomous UAV Detection

553

cles (uav) to nuclear facilities. Security Journal, pages

1–20.

Unlu, E., Zenou, E., and Riviere, N. (2017). Ordered mini-

mum distance bag-of-words approach for aerial object

identiﬁcation. In Advanced Video and Signal Based

Surveillance (AVSS), 2017 14th IEEE International

Conference on, pages 1–6. IEEE.

Villasenor, J. (2013). Observations from above: unmanned

aircraft systems and privacy. Harv. JL & Pub. Pol’y,

36:457.

Wild, G., Murray, J., and Baxter, G. (2016). Exploring civil

drone accidents and incidents to help prevent potential

air disasters. Aerospace, 3(3):22.

Yoon, J. H., Xu, H., and Carrillo, L. R. G. (2017). Ad-

vanced doppler radar physiological sensing technique

for drone detection. In SPIE Defense+ Security, pages

101880S–101880S. International Society for Optics

and Photonics.

Zhang, D. and Lu, G. (2002). Shape-based image retrieval

using generic fourier descriptor. Signal Processing:

Image Communication, 17(10):825–848.

ICPRAM 2018 - 7th International Conference on Pattern Recognition Applications and Methods

554