Sistema de detecção de crianças em situações de perigo em piscinas usando DEEP LEARNING e IOT

Abstract

This work presents a proposal to identify children in dangerous situations, in other words, alone near or inside swimming pools using Deep Learning techniques. Currently, the second leading cause of death for children in Brazil is due to drowning, in public swimming pools a lifeguard is usually used to take care of many people, and in residential swimming pools it is necessary that responsibles are fully watching children nearby. Thus, the reason of this work is to create a low-cost system capable of providing greater safety in private pools, in order to prevent possible accidents. Initially, a representative database of children and adults was consolidated, to be feed to a classification and detection models in order to evaluate them by metrics and choose the best approach. The results obtained show that the best approach was the detection with YOLOv4-tiny obtaining a mAP of 89.5%, and it will be loaded on a Raspberry Pi. The solution was implemented with a camera, an alarm circuit and messaging sending through MQTT, where the model was able to do the detections through the collected frames and validate that the child was in danger, in other words, without the presence of a adult, thus triggering the alarm and sending messages.