WILD ANIMALS CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORKS

This work aims at evaluating the use of deep convolutional neural networks (CNNs) for the classification of wild animals captured by trap cameras. Such classification is relevant in forestry management because it is an important tool for monitoring the quantity and diversity of the wild animal population in order to guarantee its long term survival and wellbeing. To train and test the CNNs, we built and made public a dataset of 5,120 images of three classes of mammal speciesof interest: armadillo, deer and wild pig. Because of the typical conditions of use of trap cameras employed in tropical forests (our use case), the resulting dataset images are very challenging to classify due to occlusions, illumination conditions (day and night), weather, etc. In our evaluation experiments, we examined three types of CNNs architectures with different configurations.Our best CNN architecture was able to correctly classify 92.80% of the images.