The world’s elderly population is expected to grow by more than triple by 2050. This indicates that detecting human activity is needed to prevent emergency situations for people living alone. Sound is an excellent resource because it has enough information to detect events and it is easy to gather. However, variability is one of the main challenges in research related to sound. To solve variability, most research focuses on selecting sound features because researches related to sound have their own purpose and suitable sound features are different for each research. In this research, we focus on classifiers to solve problem of variability. The Double-layer classification (DLC) is composed with Support Vector Machines (SVM) and the Viterbi search and detects sound events using the Hidden Markov Model (HMM). As a result, unusual sounds which occur at home such as a baby crying, a scream, a breaking glass, and a gunshot are classified by the DLC, which accuracy is 94.4%. ⓒ 2014 DGIST
Table Of Contents
I. INTRODUCTION 1-- II. RELATED WORK 6-- III. SYSTEM OVERVIEW – DOUBLE LAYER CLASSIFICATION 8-- 3.1. HMM (Hidden Markov Model) 9-- 3.1.1. Example of HMM 9-- 3.1.2. Define HMM Parameters 10-- 3.1.3. HMM for the Sound Event Detection 10-- 3.1.4. Definition of Observations 12-- 3.2. Preprocessing and Feature Extraction 13-- 3.3. First Layer SVM 14-- 3.4. Second Layer Viterbi Search 15-- 3.5. Decision Making 17-- IV. EVALUATION 19-- 4.1. Performance Comparison 19-- 4.2. Accuracy of Sound Features 20-- 4.3. Relationship between the Early Detection and Computation Time 21-- 4.4. Result of Classification Four Kinds of Unusual Sound Event 23-- 4.5. Accuracy of Frame Duration and Computation Time 23-- 4.6. Classification in a Noisy Environment 25-- V. CONCLUSIONS 28-- REFERENCES 29-- 요약문 33