Desert bighorn sheep (Ovis canadensis) recognition from camera traps based on learned features

Abstract

Monitoring wildlife in geographical areas is essential for the conservation of the biological heritage. At present, the strenuous task of animal observation on-site has been mitigated with the use of camera traps. The information gathered by these devices comprises a sequence of images, which are triggered by motion or heat, and enables the monitoring of several geographical areas at the same time without perturbing the fauna. Notwithstanding the advances of these kind of monitoring approaches, the captured images still must be manually classified, becoming into an expensive process. This paper describes an automatic methodology for labeling images captured with camera traps as a support tool for animal behaviour analysis. Specifically, we focus on the analysis of the Ovis canadensis better known as desert bighorn sheep, a species that inhabits in northwestern Mexico, USA and Canada. The importance of this species lies in that it is an emblematic one, of great historical, cultural and social value. We adopted a methodology based on a residual neural network (ResNet) as feature extractor and standard models for the classification of images depicting species of interest. The method is built (trained) and evaluated on realistic images captured by camera traps in the field. We achieve classification performances ranging from 89% for a multiclass classification setting (7 classes associated to the animal of interest) to 99% in a binary classification scenario (presence vs. absence of the species). The collected data set, model and extracted features are publicly available under request. We foresee the released data set and the proposed solution will boost research on the analysis of this species.