Abstract:
The work in this thesis presents a system to detect distress situation from sound inside home environment for elderly people. Elderly people live independently in most of the countries. The reasons may include illness and other medical conditions. They prefer to live inside their homes rather than moving to an elder care unit. In a distress situation, it becomes even more difficult for them to reach a device or contact for help. Sound carries a huge amount of information about the elderly and their surroundings. Aged persons in distress situation can make different sounds like crying or glass breaking etc. A recognition system can detect and classify these sounds.
We present the literature related to sound recognition and elder care. We also explain and compare general methods that are present in literature. Our contribution consists of collecting a set of acoustic features for daily sound classification that can effectively recognize sounds inside home environment. We compare our sound recognition results with the literature under the same experimental setup on the benchmark Real World Computing Partnership Sound Scene Database in Real Acoustical Environments (RWCP-DB) and the Sound Dataset. We create and evaluate the performance of our nine combinations of ensemble methods on these datasets. Our experimental methodology consists of first applying individual classifiers with acoustic features to measure their performance in recognizing daily sounds. We then employ ensemble methods to better recognize the daily sounds. The ensemble methods yield higher sound recognition rates as compared to the rates produced by individual classifiers. The classification performance of our proposed setup of ensemble methods shows a significant improvement in classification accuracy as compared to previous results from literature for the two datasets.
