Automated detection of European wild mammal species in camera trap images with an existing and pre-trained computer visi
- PDF / 2,189,094 Bytes
- 7 Pages / 595.276 x 790.866 pts Page_size
- 42 Downloads / 161 Views
METHODS PAPER
Automated detection of European wild mammal species in camera trap images with an existing and pre-trained computer vision model Christin Carl 1
&
Fiona Schönfeld 1 & Ingolf Profft 2 & Alisa Klamm 3 & Dirk Landgraf 1
Received: 10 September 2019 / Revised: 27 May 2020 / Accepted: 8 July 2020 # Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract The use of camera traps is a nonintrusive monitoring method to obtain valuable information about the appearance and behavior of wild animals. However, each study generates thousands of pictures and extracting information remains mostly an expensive, time-consuming manual task. Nevertheless, image recognition and analyzing technologies combined with machine learning algorithms, particularly deep learning models, improve and speed up the analysis process. Therefore, we tested the usability of a pre-trained deep learning model available on the TensorFlow hub–FasterRCNN+InceptionResNet V2 network applied to images of ten different European wild mammal species such as wild boar (Sus scrofa), roe deer (Capreolus capreolus), or red fox (Vulpes vulpes) in color as well as black and white infrared images. We found that the detection rate of the correct region of interest (region of the animal) was 94%. The classification accuracy was 71% for the correct species’ name as mammals and 93% for the correct species or higher taxonomic ranks such as “carnivore” as order. In 7% of cases, the classification was incorrect as the wrong species’ name was classified. In this technical note, we have shown the potential of an existing and pre-trained image classification model for wildlife animal detection, classification, and analysis. A specific training of the model on European wild mammal species could further increase the detection and classification accuracy of the models. Analysis of camera trap images could thus become considerably faster, less expensive, and more efficient. Keywords Computer vision . Image analysis . Camera trap . Pre-trained model . Wild mammal species
Introduction
* Christin Carl [email protected] Fiona Schönfeld [email protected] Ingolf Profft [email protected] Alisa Klamm [email protected] Dirk Landgraf [email protected] 1
Forestry and Ecosystem Management, University of Applied Sciences Erfurt, Leipziger Straße 77, 99085 Erfurt, Germany
2
Forstliches Forschungs- und Kompetenzzentrum, ThüringenForst AöR, Jägerstraße 1, 99867 Gotha, Germany
3
Nationalparkverwaltung Hainich, Bei der Marktkirche 9, 99947 Bad Langensalza, Germany
Camera traps have become an essential tool for biodiversity monitoring and wildlife management as it is a non-intrusive low-cost survey method (Newey et al. 2015) to study the occurrence and behavior of individual animals, to estimate the population’s sizes, to monitor the distribution of species (Silveira et al. 2003), and to evaluate spatiotemporal behavior (Bowkett et al. 2008). Modern units can be left unattended for a large period of time and h
Data Loading...
