Abstract
Prompt emergency detection and response in indoor environments is a significant issue due to the difficulties in detecting indoor emergency events. However, current indoor monitoring tasks are mainly carried out by manual observations of occupants and such human-dependent methods generally have limitations in taking actions against emergency events. Many researchers have made much effort to develop automated indoor monitoring systems using wearable sensing device technologies and computer vision. While these methods have various advantages, there still remain challenges to be addressed for detecting indoor emergency events; for instance, wearable sensors need to be attached to a human body and occlusions make it hard to recognize the emergencies. To overcome those deficiencies, this paper proposes a sound event recognition (SER)-based indoor event classification (e.g., emergency and normal event) method with a convolutional neural network (CNN). The research consists of four main steps. First, the sound types of indoor events are determined as four emergency sounds (explosion, gunshot, glass break, and scream) and one normal sound (sleeping). Second, 692 sound data of identified events are collected from online sound data sharing services, and the preprocessing is performed. Third, SER model is developed through CNN algorithm with log-scaled mel-spectrogram features. Finally, model performance is evaluated using 5-fold cross validation. The experimental results showed that the sounds caused by indoor emergency events could be automatically recognized by the proposed method with F-score of 77.32%, which demonstrates its applicability for real emergency situations.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call