R-count: Ferramenta de contagem remota de bovinos

Abstract

The performance of accurate and reliable detections of animals by the artifice of digital aerial images (IDAs) obtained by Remotely Piloted Aircraft (RPAs) shows itself as a very promising area. However, extremely challenging in implementation for the remote and intelligent management of animals for the near future. In this way, this work has as main intention the applicability of a model of Deep Convolutional Neural Network (DCNN) in a context that approaches the counting of cattle both for the context of pastures and confinement. For this, in the first stage, imaging was performed via RPA of 125 F1 Angus-Nelore females, divided into two batches in the final fattening phase. In the second stage, at the U. Farm - Smart Collab of the Institute of Animal Health and Production of the Federal Rural University of the Amazon (UFRA), all the activities of pre-processing of images and training of Neural Networks took place. In this way, two databases were prepared, respectively divided into training and testing, in proportions (70:30). The results from the Mask R-CNN with ResNet101 for DAU (Database Augumentation) after 100 epochs and 179 minutes of training were more promising when comparing MAask R-CNN with ResNet50 for DAU after 100 epochs and 140 minutes of training, obtaining average accuracy mAP of approximately 95%, having the best detections of animals on pasture, consequently providing opportunities for counts.