CATTLE SEGMENTATION USING UAVS SUPPORTED BY COMPUTER VISION TECHNIQUES: PRELIMINARY RESULTS

Abstract

We can list some applications of systems based on interpretations of images captured by UAVs (Unmanned Aerial Vehicle) as support for precision livestock farming. Among them we highlight the counting of cattle heads for inventory and other purposes, tracking, identification and slaughter point estimation. These applications can be supported by computer vision techniques, however, some approaches work with the entire captured image without separating the object of interest from other objects in the image, hence processing noise leading to undesired performance. Separating objects of interest into images can improve the performance of applications, since errors caused by noise are discarded. In this sense, we propose an approach to segment cattle in feedlot batches , through images acquired by UAVs using computer vision and machine learning techniques. For this, we used aerial images captured with a DJI Phantom 4, the height of 20 meters in a flight that resulted in a set of 101 images of an experimental confinement installed in the Fazenda Escola of UFMS/FAMEZ. The animals were arranged in individual paddocks with a structure equipped with feeders and water through , totaling 22 Nellore cattle.

Full text

GRSS-YP & ISPRS Student Consortium SS 2018 Campo Grande, 2018. From October 29

th

to November 1

st
CATTLE SEGMENTATION USING UAVS SUPPORTED BY COMPUTER
VISION TECHNIQUES: PRELIMINARY RESULTS
VANESSA APARECIDA DE MORAES WEBER 2,3
FABRICIO DE LIMA WEBER 1,3
ADAIR DA SILVA OLIVEIRA JUNIOR 1
GILBERTO ASTOLFI 1,4
MARIA ISTELA CAGNIN MACHADO 1
MARINA DE NADAI BONIN 1
LUIS CARLOS VINHAS ÍTAVO 1
EDUARDO SOUZA LEAL 1
JOSÉ MARCATO JUNIOR 1
HEMERSON PISTORI 1,2
1 FEDERAL UNIVERSITY OF MATO GROSSO DO SUL (UFMS)
CAMPO GRANDE - MS
adairsojr@gmail.com; istela@facom.ufms.br;  marina.bonin@ufms.br; luis.itavo@ufms.br;
eduardoleal.zoo@gmail.com; jose.marcato@ufms.br
2 DOM BOSCO CATHOLIC UNIVERSITY (UCDB)
CAMPO GRANDE - MS
vamoraes@gmail.com; pistori@ucdb.br
3 STATE UNIVERSITY OF MATO GROSSO DO SUL (UEMS)
MASTER IN ANIMAL PRODUCTION IN THE CERRADO-PANTANAL, AQUIDAUANA - MS
fabricio.weber@gmail.com;
4 FEDERAL INSTITUTE OF EDUCATION, SCIENCE AND TECHNOLOGY OF MATO GROSSO DO SUL (IFMS)
CAMPO GRANDE - MS
gilberto.astolfi@ifms.edu.br
We can list some applications of systems based on interpretations of images captured by UAVs (Unmanned
Aerial Vehicle) as support for precision livestock farming. Among them we highlight the counting of cattle
heads for inventory and other purposes, tracking, identification and slaughter point estimation. These
applications can be supported by computer vision techniques, however, some approaches work with the
entire captured image without separating the object of interest from other objects in the image, hence
processing noise leading to undesired performance. Separating objects of interest into images can improve
the performance of applications, since errors caused by noise are discarded. In this sense, we propose an
approach to segment cattle in feedlot batches, through images acquired by UAVs using computer vision and
machine learning techniques. For this, we used aerial images captured with a DJI Phantom 4, the height of
20 meters in a flight that resulted in a set of 101 images of an experimental confinement installed in the
Fazenda Escola of UFMS/FAMEZ. The animals were arranged in individual paddocks with a structure
equipped with feeders and water through, totaling 22 Nellore cattle. We segment captured images using
the Simple Linear Iterative Clustering (SLIC) algorithm, configured to segment the image into 850 segments,
44

GRSS-YP & ISPRS Student Consortium SS 2018 Campo Grande, 2018. From October 29

th

to November 1

st
Gaussian kernel diameter at 5.0 and compactness, which balances the proximity of color and space to 10.
The pixels grouped by the SLIC formed superpixels totaling 202 segments. We visually classify superpixel
segments of an image into two classes: bovine and background. As background were considered fences,
troughs, drinking fountain, land and grass regions. After classifying the superpixel segments we extracted
the characteristics of each of them using the descriptors of Histogram of Oriented Gradients (HOG), Hu's
Moment Invariants, Gabor Filter (GFs), Local Binary Patterns (LBP), Gray Level Co-occurrence Matrix (GLCM)
and information in the RGB, HSV and CIELAB color spaces. The characteristics extracted from the
superpixels were submitted as input to the K-Nearest Neighbour hood (KNN), Support Vector Machines
(SVM), Decision Trees (J48) and AdaBoost machine learning algorithms with the WEKA 3.8.2 standard
parameters . We performed training and testing on the extracted data and used the 10-fold cross-validation
to obtain the metric precision, with the same number of bovine and background images. We obtained the
highest precision with the KNN algorithms (k = 1) with 97.02%, followed by SVM and AdaBoost, both with
94.55%. The worst accuracy was achieved with the J48 algorithm with 92.57%. Although this experiment
characterizes the initial stage of segmentation of cattle in images captured with UAVs, we believe that the
presented results are encouraging and may support other researchers to develop several softwares whose
process needs to identify the bovine as main object in this type of image. In addition, stimulate new
approaches such as KNN with other K values, validate results by means of segmentation metrics among
others. Finally, to explore the use of computer vision and machine learning to segment cattle in aerial
images.
45