Aprendizado profundo aplicado a imagens aéreas para identificação de bovinos em campos de engorda

Abstract

This work addresses the use of computer vision techniques for the automatic identification of animals in feedlots, a relevant task for monitoring and management in precision livestock farming. However, there are still gaps in the literature regarding the development of specific datasets capable of reducing duplicate detections and increasing detection accuracy in real-world scenarios. Thus, the objective of this dissertation is to develop and quantitatively evaluate a dataset for cattle identification in feedlots, aiming to increase recall and reduce the occurrence of duplicate detections in aerial images. For this purpose, images collected by a remotely piloted aircraft equipped with an RGB camera were used, with the data stored on an SD card for subsequent processing. From these images, animal fragments were extracted and used to train a neural network through transfer learning, allowing weight adjustment for recognition under different positions and angles. As results, the model pre-trained on the COCO dataset achieved 90% precision, 67% recall, and an F1-score of 76.82%, indicating that although it correctly classifies detected objects, it fails to detect a significant portion of the cattle present in the images. In contrast, the JGBCowV1 dataset achieved 73% precision, 82% recall, and an F1-score of 77.24%, while the refined JGBCowV1.1 version reached 83% precision, 75% recall, and an F1-score of 78.79%, demonstrating consistent improvement in model performance and a better balance between precision and detection capability when compared to the COCO dataset, in addition to superior performance in cattle identification under the real-world conditions analyzed. Finally, a consolidated results file was generated and, as future work, the use of additional sensors, such as thermal cameras, is suggested to improve detection in scenarios with high animal density.