Análise de desempenho de Yolov3 quanto a variação de fatores da cena em imagens para detecção de personagens em vídeo

Abstract

Information is the result of processing, manipulation, and data organization in such way that can mean a modi cation in the system knowledge. In the last years, the volume of data has grown exponentially and for that also has grown the need of treating these data more e ciently, making them meaningful and trustworthy information source for those systems which need performance and a reliable response. In this context, Arti cial Intelligence (AI) appears as a tool for understanding and solving problems in other paradigm, making systems into smart ones. So, this work contemplates YoloV3, framework for a state of art technique in object detection and recognition. Its objective is to check what parameters of the images make recognition harder. This work picked three parameters to test: positioning of the subject, background, and illumination. Results showed that images which the character is looking down and towards the opposite way, or with high brightness levels or too di erent backgrounds used in the training section have impact on the network accuracy. To avoid those situations, images processing techniques should be used as well as increasing the images in the training section.