Método aplicado ao monitoramento remoto de animais vaseado em aerolevantamento com VANT e aprendizagem profunda

Abstract

One of the main demands in the extensive livestock management is the counting of animals in areas of tens of hectares, costly when performed manually and locally. In such context, this work proposes and discusses the effectiveness of a semi-autonomous, noninvasive method for remote identification of animals in the field, applicable to precision livestock farming systems. The method was conceived from an applied experimental methodology, exploratory at first, explanatory later, based on remote sensing techniques that include image collection processes by aerial surveying with a RGB camera embedded in an unmanned aerial vehicle, the persistence of the images obtained by means of storing spatiotemporal data, and processing of stored images to generate an orthomosaic map covering the rural property succeeded by the application of pattern discovery processes, making use of machine learning, especially convolutional neural networks. According to experiments carried out in a real scenario, the method proved to be effective, being able to detect and count animals with an accuracy of up to 92% from different sets of images collected using UAV at 100 m, and it is still possible to identify the position the animals to the field.