Real-Time Sound Event Classification for Human Activity of Daily Living using Deep Neural Network
Ah Hyun Yuh, Soon Ju Kang · 2021
Over the past years, increasing number of IoT sensors played important role in developing ambient assisted living (AAL) technologies such as elderly home care system by predicting activity of daily livings (ADLs). One way to develop smarter home care services with unobtrusive sensors in ubiquitous forms is using sound. This paper suggests a methodology to detect different sound events generated by residents based on real-life audio data. We propose a guide all the way from installing wireless microphone networks to recording, annotating, and preprocessing audios. Then we extract audio features and design deep learning classifier to classifying sound events. Finally, we deploy classifier on real-life scenarios to implement sound event detection in real-time. We evaluated 2D convolutional classifier with 16 sound events, achieving 95.55 % training accuracy, 94.64 % validation accuracy, 96.40 % recall score, and 94.93 % F1-score.