Tracking Animal Behavior Using Deep Convolutional Neural Networks

U okviru ovog rada proučene su duboke neuronske mreže za automatizirano praćenje životinja, točnije praćenje peradi na farmi. Testirana su dva modela kamere za dohvaćanje slika, od kojih se jedan temelji na točkastom modelu a drugi na modelu širokokutne kamere. Model Mask R-CNN korišten je za segmentaciju instanci objekata. Optimizirani su hiperparametri modela, testirani različiti izlučivači značajki i konačno dana srednja preciznost kao mjera evaluacije modela.Within this paper, deep neural networks have been studied for automated animal tracking, namely the monitoring of poultry on the farm. Two camera image retrieval models have been tested, one of which is based on a pinhole model and the other on a fisheye camera model. Model Mask R-CNN was used for segmentation of animal instances. Hyperparameters of model were optimized, different backbone architectures were tested, and finally, average precision was given as a measure of model evaluation