Drone-Based Cattle Detection Using Deep Neural Networks

Document Type

Conference Proceeding

Source of Publication

Advances in Intelligent Systems and Computing

Publication Date

1-1-2021

Abstract

© 2021, Springer Nature Switzerland AG. Cattle form an important source of farming in many countries. In literature, several attempts have been conducted to detect farm animals for different applications and purposes. However, these approaches have been based on detecting animals from images captured from ground level and most approaches use traditional machine learning approaches for their automated detection. In this modern era, Drones facilitate accessing images in challenging environments and scanning large-scale areas with minimum time, which enables many new applications to be established. Considering the fact that drones typically are flown at high altitude to facilitate coverage of large areas within a short time, the captured object size tend to be small and hence this significantly challenges the possible use of traditional machine learning algorithms for object detection. This research proposes a novel methodology to detect cattle in farms established in desert areas using Deep Neural Networks. We propose to detect animals based on a ‘group-of-animals’ concept and associated features in which different group sizes and animal density distribution are used. Two state-of-the-art Convolutional Neural Network (CNN) architectures, SSD-500 and YOLO V-3, are effectively configured, trained and used for the purpose and their performance efficiencies are compared. The results demonstrate the capability of the two generated CNN models to detect groups-of-animals in which the highest accuracy recorded was when using SSD-500 giving a F-score of 0.93, accuracy of 0.89 and mAP rate of 84.7.

ISBN

9783030551797

ISSN

2194-5365

Publisher

Springer

Volume

1250 AISC

First Page

598

Last Page

611

Disciplines

Computer Sciences

Keywords

Convolution Neural Networks, Drones, Object detection, Unmanned aerial vehicles

Scopus ID

85090178320

Indexed in Scopus

yes

Open Access

no

Share

COinS