以建築資訊模型/深度學習作法實現聲音定位救災用途;Realization of Indoor Positioning Based on Sound for Disaster Relief Using Building Information Modeling and Deep Learning

NCUIR > Engineering College > Graduate Institute of Civil Engineering > Electronic Thesis & Dissertation > Item 987654321/79337

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/79337

Title:	以建築資訊模型/深度學習作法實現聲音定位救災用途;Realization of Indoor Positioning Based on Sound for Disaster Relief Using Building Information Modeling and Deep Learning
Authors:	張智雄;Chang, Chih-Hsiung
Contributors:	土木工程學系
Keywords:	人工智慧;聲音定位;虛擬實境;建築資訊模型;頭部相關轉移函數;Artificial Intelligence;Sound Positioning;Virtual Reality;Building Information Modeling;Head Related Transfer Functions
Date:	2019-01-04
Issue Date:	2019-04-02 14:07:29 (UTC+8)
Publisher:	國立中央大學
Abstract:	隨著科技的日新月異，建築資訊模型(Building Information Modeling, BIM)、人工智慧(Artificial Intelligence, AI)和虛擬實境(Virtual Reality, VR)技術也逐漸成熟，其中又以AI的成長最為顯著，此一進展也漸漸改變人們的生活習慣，例如透過手機聲音辨識即可達成語音查詢或個人化秘書之功能。而聲音對於建築物也是一項非常重要之課題，例如透過各種科技模擬建築物室內或戶外場景中的聲音傳播和噪音控制…等。目前，聲音定位方面之應用非常具有潛力，但聲音定位技術仍會受到噪音、障礙物…等干擾，所以運用較為受限，故本研究想藉此提出一聲音定位的方法，以解決上述之問題。本研究利用BIM、VR和頭部相關轉移函數(Head Related Transfer Functions, HRTF)技術達到擬真音場，並蒐集音訊後再透過AI訓練，以達到聲音定位之功能。經過驗證我們得知：利用本研究所提出之聲音定位方法，在解析度為25m2時，準確度為95.2%；在解析度為12m2時，準確度為89.3%，且在AI訓練階段皆有加入訊噪比為0.3之噪音。在連續性音訊中，當解析度為25 m2和12 m2時，都可100%辨識，且準確度分別高達94%及86%。本研究驗證區域也挑選障礙物較多之廚房和客廳區域，相對於傳統聲音定位受障礙物之影響，透過本研究所提出之聲音參數化架構，反而於AI訓練階段取得更多特徵值，以達到更高準確度，也解決現今聲音定位應用範圍受限之疑慮，且相較於其他無線射頻定位技術受金屬屏蔽作用所限制，本研究所提供聲音定位之功能可利用蒐集不同音頻，並擷取某特定頻率進行定位，以此增加在複雜環境聲音定位可行性，例如：火場…等。;With the rapid development of science and technology, BIM, AI, and VR technologies have gradually matured. Among them, the growth of Artificial Intelligence has become most remarkable. This progress has gradually changed people’s living habits. For example, a voice query or personalized secretary function can be achieved through voice recognition of a cell phone. Sound is also a very important issue for buildings, such as simulating the sound transmission and noise control in indoor or outdoor scenes of buildings through various technologies. At present, the application of sound positioning has great potential, but sound positioning technology is still subject to noise, obstacles, etc. Therefore, this study would like to propose a sound localization method to solve the above problem. This study uses BIM, VR, and HRTF techniques to achieve a Virtual sound field. Collecting the audio and then uses AI training to achieve sound positioning. After verification, we learned that: Using the sound positioning method proposed in this study, in the AI training stage, noises with a signal to noise ratio 0.3 are added. The accuracy is 95.2% when the resolution is 25m2; The accuracy is 89.3% when the resolution is 12m2. In continuous stream of audio, when the resolution is 25 m2 and 12m2, it can be 100% recognized, and the accuracy is as high as 94% and 86% respectively. In this study, the verification area selected kitchen and living areas with more obstacles. Compared with traditional sound localization affected by obstacles, through the sound parameterization framework proposed in this study, more features are obtained in the AI training stage to achieve higher accuracy. Also solves the doubts about the limited scope of today′s sound positioning applications. Compared to other radio frequency positioning technologies, it is limited by metal shielding. The sound positioning function of the institute can be used to collect different frequencies and capture a specific frequency for positioning. In order to increase the feasibility of sound positioning in a complex environment, such as fire field, and so on.
Appears in Collections:	[Graduate Institute of Civil Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	188	View/Open

社群 sharing

Loading...