A Framework for IoT-Based Audio Recognition of Crying Children to Prevent Trap in a School Van.

This paper presents a framework to detect crying children left unattended in a car during the daytime. The detection triggers an alert to rescuers, parents, the driver, and any persons in charge. The solution contained two parts: hardware with the raspberry pi and software implemented with the deep-learning technique. In the model training phase, a sound dataset was collected online (crying children, ambient sound, and people talking). The sound was transformed into an MFCC representation to change the audio recognition into an image representation. One pre-trained convolutional neural network (CNN) model was used for the classification. It classifies sounds into three classes: a crying baby, people speaking, and other sounds different from the two previous ones, such as dog barking, noise, and sirens). The detection accuracy is 78%, with sounds recorded by a microphone in real-time.