Animal recognition using deep learning

Abstract

Camera traps are widely used for wildlife monitoring. In this work machine learning based data processing pipeline is assembled for animal detection on the camera trap images focusing on the ungulate species. The typical animal detection challenges are noted, and available solutions are evaluated. As the result of this work, two different deep neural networks Faster R-CNN and RetinaNet were trained, achieving 0.2786 mAP@0.5:0.05:0.95 and 0.4562 mAP@0.5 on the dataset of interest gathered in the Latvian forest regions during the ”ICT-based wild animal census approach for sustainable wildlife management” project. Additionally, different learning optimization techniques such as data augmentation and oversampling were implemented and assessed.