Utilização da análise quantitativa de recorrência e coeficientes cepstrais de frequência mel para reconhecimento de incêndio em vegetação usando sinal de áudio

In this work it was sought to analyze the possibility of detecting wildfire through audio signals. Therefore, it was idealized the usage of a learning algorithm, more especifically a Support Vector Machine, that performs the task of binary classification between audios corresponding or not to wildfire according to the features that feed him. To train this machine were obtained on the internet audios of wildfire and typical sounds of forest and field environments like sounds of animals, wind, running water and rain for exemple. The main technique proposed to extract the features from the signals is the Recurrence Quantification Analisys which is mainly designed for the analisys of chaotic or nonlinear systens. Allied to that, it was employed the Mel-Frequency Cepstral Coefficients which made the application of that technique easier and also contributed increasing the accuracy of the classification algorithm that, in the best trained models, achieved 95.1% of accuracy and a false negative rate of 3%. Thus, it was concluded that the utilized methods are satisfactorily able to detect wildfire in audio signals when compared to the results exposed in other papers.