Artificial intelligence monitoring at the edge for smart nation deployment

• المؤلف الرئيسي: Leow, Hui Wen
• مؤلفون آخرون: Gan Woon Seng
• التنسيق: Final Year Project
• اللغة: English
• منشور في: Nanyang Technological University 2021
• الموضوعات: Engineering::Electrical and electronic engineering
• الوصول للمادة أونلاين: https://hdl.handle.net/10356/150106
• المؤسسة: Nanyang Technological University

Environmental noise pollution has been a key challenge faced by large cities in recent days. Concerned with its detrimental impacts on the well-being of people, smart nations today are seeking data-driven solutions to combat noise pollution. Sound Event Recognition (SED) systems, which are able to identify sources of unwanted noise have since therefore grown in demand and research interest. To support research efforts on audio classification of sounds, this paper thus presents the SG Soundscapes audio dataset developed by a team of 5 members. While open-source audio datasets are already available in the research community, this dataset in particular adopts a strong multi-class multi-label labelling approach on urban sounds collected in Singapore. This means that the event label, onset and offset as well as proximity of sounds are annotated. Annotations were not crowdsourced and are wholly calibrated among members in the team. The taxonomy is also collaboratively designed, labelling methods were agreed upon and audio devices used for listening were standardized. Furthermore, research findings investigating the performance accuracy of a deep learning model trained using datasets combined from geographically independent spaces are also presented. Results have shown that domain mismatch, varying ambient noise as well as different annotation methods are key factors that could affect model performance.