An Unobtrusive Event Recognition System for Smart Homes Using Audio Signals
Title: An Unobtrusive Event Recognition System for Smart Homes Using Audio Signals
Authors: Anastasios Vafeiadis, Konstantinos Votis, Dimitrios Giakoumis, Dimitrios Tzovaras, Liming Chen and Raouf Hamzaoui
Abstract—Internet of Things (IoT) devices are generating a large amount of data in different forms such as, text, audio, video, biometric, etc. Building an acoustic-based event recognition system for Smart Homes (SH) is a challenging task due to the lack of high-level structures in environmental sounds. Furthermore, the audio data collection is a particularly difficult process, since it requires a proper equipment in capturing the audio signals. Labeling the collected data is an essential step for the recognition system, since overlapping events can occur. In particular, the selection of effective features is still an open problem. We make an important step toward this goal by showing that the combination of Mel-Frequency Cepstral Coefficients, Gammatone Frequency Cepstral Coefficients, Spectral Roll-off, Spectral Centroid Zero-Crossing Rate, and Discrete Wavelet Transform features can achieve an F1 score of 90.2% and a recognition accuracy of 91.7% with a gradient boosting classifier for ambient sounds recorded in different kitchen environments and an F1 score of 96.7% and a recognition accuracy of 98.0% with a Convolutional Neural Network (CNN).
IEEE Transactions on Big Data Special Issue on Big Data in Ubiquitous Computing
First Round Decisions: December 15th, 2017 (PAPER IS UNDER SUBMISSION)