Reconhecimento de fauna silvestre com o uso de deep learning

Abstract

Automated wildlife monitoring has gained prominence in environmental studies, particularly with the advancement of computer vision techniques. This work presents the development and validation of an automated wildlife detection system using convolutional neural networks, specifically the YOLOV8 architecture, applied to footage from camera traps. The methodology involved processing 60 videos originally collected between October and November 2021, resulting in 19.808 frames manually annotated using the CVAT tool. The dataset encompassed seven species of Brazilian fauna: Sus scrofa (wild boar) 61.7% of occurrences, Cuniculus paca (paca) 10%, Nasua nasua (coati) 8.3%, Salvator merianae (teiú lizard) 8.3%, Dasypus novemcinctus (nine-banded armadillo) 6.7%, Didelphis (opossum) 5% and birds (generalized class) 3.3%. Training was conducted in progressive stages, starting with 3,000 images and gradually expanding to the full dataset, using accessible hardware (GTX 1060 GPU) with a total processing time of 50 hours and 25 minutes. The results demonstrated the operational feasibility of the model, showing accurate detection of the target species even under adverse conditions of nighttime illumination and partial vegetation occlusion. Performance became more consistent with an increased amount of data, with particular effectiveness in identifying Sus scrofa (mAP50 of 0.734), an invasive exotic species of relevance to environmental control policies