Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 92 (2016) 13 17

2nd International Conference on Intelligent Computing, Communication & Convergence

(ICCC-2016)
Srikanta Patnaik, Editor in Chief
Conference Organized by Interscience Institute of Management and Technology
Bhubaneswar, Odisha, India

Motion Based Animal Detection in Aerial Videos

Yunfei Fanga*, Shengzhi Dub, Rishaad Abdoolaa, Karim Djouania, Coneth Richardsa
a

Electrical Engineering, Faculty of Engineering and the Built Environment

Tshwane University of Technology, Pretoria 0001, South Africa
b
Mechanical Engineering, Mechatronics and Industrial Design, Faculty of Engineering and the Built Environment
Tshwane University of Technology, Pretoria 0001, South Africa

Abstract
Computer vision techniques are applied to perform automatic wildlife surveying and animal monitoring. Animal detection in
aerial videos is challenging because of the complexity of wild environments. In this paper, a method for moving animal detection
is proposed by taking advantage of global patterns of pixel motion. In the video dataset, where animals make obvious movement
against the background, motion vectors of each pixel are estimated by applying optical flow methods. A coarse segmentation
then removes most parts of the background by applying a pixel velocity threshold. Based on the segmented regions, another
threshold was employed to filter out negative candidates that could belong to the background. The pros and cons of this method
are discussed.
2014
TheAuthors.
Authors.Published
Published
Elsevier
2016 The
byby
Elsevier
B.V.B.V.
This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Selection
and peer-review under responsibility of scientific committee of Missouri University of Science and Technology.
Peer-review under responsibility of the Organizing Committee of ICCC 2016
Keywords: optical flow; image segmentation; computer vision; object detection; animal tracking; dynamic background; moving camera; UAV;
drone; wildlife survey;

* Corresponding author. Tel.: +27-(0)74-586-2926..

E-mail address:fangyunfei08@gmail.com

1877-0509 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the Organizing Committee of ICCC 2016
doi:10.1016/j.procs.2016.07.316

14

Yunfei Fang et al. / Procedia Computer Science 92 (2016) 13 17

1. Introduction
In nature conservation management, wildlife surveys are important either to provide animal population and
abundance information or to monitor threats to animals. Such surveys are usually conducted on the ground by
conservation workers. It is labour intensive with a high financial cost. Aerial videos of wildlife collected by UAVs
(unmanned aerial vehicles) provide a great substitution if combined with proper automatic detection techniques.
Object detection and tracking has been developed for decades within the field of computer vision. However, all
algorithms proposed have restrictions based on the specific circumstances. Moving object detection in stationary
cameras with a constant background can be easily handled today. A great amount of trackers have been formed utilizing
different methods, such as background modelling methods [1] and State Estimation methods [2]. Addressed the problem
of moving object detection from moving cameras, a typical method is the extension of background subtraction [3] [4].
In [5], a panorama of the video sequences was generated by applying a geographic information system (GIS), and the
background was modelled by registration methods.
However, animal detection from aerial videos is beyond the reach of such algorithms. The context we are set to is
totally different from the conventional applications such as aerial traffic management [6] [7], civil surveillance [8] [9],
and military operations [10]. Observed above, animals dont have a stable contour, and consistent movement orientation
and speed, which are usually assumed in vehicle detection. Besides, the dynamic background and natural camouflage of
the animals will bring more challenges.
In this paper, an animal detection method is proposed by highlighting the different motion patterns between the
background and the animals. The advantage of motion feature is that it takes little account on the texture, colour, and
influence of illumination which form the major difficulties in our context. Experiments are done on aerial videos
captured from a real world wild scene, and the results are discussed in detail.
2. Method design
2.1. Optical flow theory
Optical flow illustrates the pattern of apparent motion between the observer and objects in visual scene. The HornSchunck algorithm [11] assumes smoothness in the flow over the whole image, which is denoted as,

(1)
The flow is formulated as a global energy functional. Solving the associated multi-dimensional Euler-Lagrange
equations, the energy functional can be minimized to obtain the optical flow of each pixel.
2.2. Animal detection
Fig. 1 illustrates the flowchart of the algorithm. Fig.2 (a) shows the original frame of a video sequence captured from
a UAV. Several zebras in the scene move in different directions while the UAV tries to follow the herd. From the
optical flow vectors shown in Fig.2 (b), it is possible to recognize the moving zebras based on the relative motion
difference between the foreground and background. Because the background motion is caused by the undertaking of the
observer, optical flow vectors of the background pixels exhibit little deviation in exterior orientation and motion
magnitude.

Video
Frames

Draw
Boundingboxes

Velocity
Threshold
(Threshold-I)

Compute
Optical Flow

Blob
Threshold
(Threshold-II)

Morphological Operation
& Blob Analysis

Fig. 1. Flowchart of the algorithm

Yunfei Fang et al. / Procedia Computer Science 92 (2016) 13 17

15

As velocities of foreground pixels generally have greater values, a threshold of global mean value of pixel velocities
is set to filter out potential background pixels. This step is referred to as Threshold-I, in which pixel values are set to

255 when the velocity exceeds the threshold, and 0 when not. A binary mask is formed and a blob test is required to
Fig. 2. (a) Original video frame (b) Optical flow vectors (c) Binary mask without morphologic operation (d) Binary mask after morphologic operation (e)
Segmentation with great errors (f) Detection result

segment the connected regions. Before the blob test, morphologic operations are required to remove speckle noise and
fill holes in the blobs, as shown in Fig. 2 (c) and (d).
Due to optical parallax, motion velocities of pixels close to the camera exhibit greater changes than those far from
the camera, which makes Threshold-I not effective to all the background ground pixels. In other cases, noises are
introduced by the influence of illumination. These will make the mask very chaotic in certain situations, as shown in
Fig.2 (e). A threshold is imposed on the blobs by calculating the statistical mean value of the largest velocities with 30%
confidence interval of each region. This step will be referred to as Threshold-II. The largest velocities extract the motion
difference between the foreground and background, and the mean value will disregard the area of the regions. This is
effective to the cases described above, of either large regions of background pixels or few noise pixels with great
velocities. Since the prior knowledge about the number of animals in view is not available, it is better to choose an
empirical value based on the motion of the animal. A bounding box will be drawn on the detected regions to form the
final result, as shown in Fig. 2 (f).
3. Experiment
Based on the assumption that animals make obvious movement against the background, foreground pixels can be
distinguished by the motion velocity.

16

Yunfei Fang et al. / Procedia Computer Science 92 (2016) 13 17

(2)
where V is the foreground pixel motion velocity, is the average of all the pixels motion velocity, coefficient is
an empirical value to select foreground pixels more efficiently. The smaller is, the more pixels are defined as potential
foreground. is decided by the motion difference level between foreground and background, and also by the accuracy
of optical flow estimation.
This algorithm was tested on two short videos sequences with two animal species that are zebras and antelope
respectively. Fig.3 shows some of the detection results in subsequent frames. The performance achieved is shown in
Table 1. Animals close to the camera composed of more pixels achieve better accuracy. The relatively higher false
positive rate in the footage of antelope is caused by the poor optical flow estimation.
Table 1. Detection performance

species

Total
Frames

Ground
Truth

True
Detection

False
alarm

False
positive

False
Negative

zebras

448

4102

3365

96

2.03%

17.97%

antelope

145

992

879

124

12.50%

11.39%

Fig. 3. Detection results

This method is sensitive to animal movement and takes little account on other features such as the size, colour and
shape appearance of the animal. Fig.4 shows the tracking result of a zebra making a sudden stop during the running and
walking in the opposite direction. It is successfully detected because of different movements between the zebra and the
background.
As shown in Fig.5, animals easily detected as one. In some cases, an animal was divided into two parts, because
some pixels of the body were filtered in Threshold-I. It is supposed that further processing, using the texture and colour
space segmentation of animals, will be helpful for this situation.

Fig. 4. Detected zebra in subsequent frames

Fig. 5. Two zebras detected as one and one divided into two parts

Yunfei Fang et al. / Procedia Computer Science 92 (2016) 13 17

4. Conclusion
Herein we investigate the possibility of combing UAVs with computer vision techniques to benefit the field of nature
conservation. An investigation into the complexity of animal detection and tracking in a wild scenario was completed.
Great challenges are introduced by the moving camera, dynamic background and aerial perspective of the scenes. A
motion based detection method was proposed by highlighting global pixel motion difference between the animal and the
background. Experiments were done using footage captured from representative environment of the wild. The results
were discussed in detail. This method showed effectiveness in the scenario of animal detection and tracking in the wild.
It highlights the efficiency of motion features in tracking. However, it cannot be assumed that animals will always be in
motion in the wild. Future work should be considered by combining texture and colour space segmentation for
stationary animals and more effective local threshold selection methods.
References
1.

C.Stauffer and W.E.L. Grimson. Adaptive background mixture models for real-time tracking,Proc. CVPR. 1999.

2.

Bhat, K.S., Saptharishi, M., Khosla, P.K,Motion detection and segmentation using image mosics,IEEE ICME, 2000.

3.

V.Reilly, H, Idrees and M. Shah, Detection and tracking of large number of targets in wide area surveillance, Computer Vision ECCV,
pp.186-199, 2010.

4.

Y. Wang, Z. Zhang and Y. Wang, Moving Object Detection in Aerial Video, 11th Inter-national Conference on Machine Learning and
Applications, 2012, pp: 446-450.

5.

Lin, C., Pankanti, S., Ashour, G., Porat, D., Smith, J.R., Moving camera analytics: Emerging scenarios, challenges, and applications, IBM
Journal of Research and Devel-opment, 2015, Vol 59, pp: 5:1-5:10.

6.

Hailing Zhou, Hui Kong, Lei Wei, Creighton, D., Efficient Road Detection and Tracking for Unmanned Aerial Vehicle, Transactions on
Intelligent Transportation Systems, 2015, Vol 16, pp: 297-309.

7.

Moranduzzo, T., Melgani, F., Automatic Car Counting Method for Unmanned Aerial Ve-hicle Images, Geoscience and Remote Sensing,
2014, Vol 52, pp: 1635 1647.

8.

Stephen Se, Feng Shi, Xin Liu, Mohen Ghazel, Object and activity detection from aerial video, Geospatial Informatics, Fusion, and Motion
Video Analytics, 2015.

9.

Prokaj, J., Medioni, G., Persistent Tracking for Wide Area Aerial Surveillance, Computer Vision and Pattern Recogniton, 2014, pp: 11861193.

10. S. Parameswaran, C. Lane, B, Bagnall, H. Buck, Marine Object Detection in UAV full-motion video, Airborne Intelligence, surveillance,
Reconnaissance Systems and Applica-tions, 2014.
11. Berthold K. P. Horn, Brian G. Schunck, Determining optical flow, Techniques and Ap-plications of Image understanding, 1981, Vol 281,
pp: 319-331.

17