Implementación de un prototipo de emisión de alertas en zonas de peligro para el cuidado de niños de 1 a 5 años mediante visión artificial.

Abstract

The present degree work encompasses all the artificial vision criteria, opting for a technology that allows this process to be carried out in real time within certain areas such as danger zones, in each zone the objects that need to be detected are different and for this reason a technology that can be trained in a personalized way was selected. Yolo v5 technology was used for object detection due to its ability to detect in real time and at high speed. For its part, pytorch in training generates deep learning in the last network of neurons. Based on the personalized training for each danger zone, a confusion matrix was obtained that shows the percentage of false positives when an object was detected without being present in the validation scenario. Therefore, the highest error percentage in the first danger zone corresponds to the fork object with a confusion of 47% and in the second danger zone to the Child class with a value of 42% confusion. In the test scenarios the system showed good results detecting each object quickly and accurately, thanks to this it was possible to determine the exact position of each detected element in order to calculate the distance between the object of interest and the child. If the distance between the Child class and the object of danger is less than 300 pixels corresponding to one meter away, the tutor is notified by two different methods, an alert issued by message through the Telegram platform with a delay of 1 second and another audible alert with a 2 second delay.