林政宏Lin, Cheng-Hung張仲軒Chang, Michael-Austin2025-12-092025-08-062025https://etds.lib.ntnu.edu.tw/thesis/detail/f55e62ae5c0036a8bea59cb0bcfe5b6d/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/125054隨著人工智慧與深度學習技術的快速發展，人體動作辨識在醫療照護、監控系統、人機互動等領域展現出極高的應用潛力。然而，傳統的影像辨識技術多仰賴可見光或紅外線攝影機，不僅容易受到環境光源變化影響，可能還有潛在的隱私疑慮。為解決上述問題，本研究提出一種結合超聲波訊號與深度學習之手部動作辨識模型，利用聲波反射特性來辨識手部動作，從而克服光線限制並提升隱私保護性。本研究使用USB介面的超聲波收音設備，錄製人體動作引發的聲波變化，並透過短時距傅立葉轉換（STFT）將訊號轉換為頻譜圖，以提取含有時頻解析度的特徵，接著使用ResNet-50卷積神經網路（CNN）進行手部動作分類。為驗證本方法之效能，本研究建立了一個包含五種手部動作及一類靜態背景的超聲波資料庫，並透過多位受試者進行測試以評估模型效能。實驗結果顯示，在特定實驗環境下，模型的辨識準確率可達95%；即使在不同受試者的推論測試中，仍能維持92%的表現。With the rapid advancement of artificial intelligence and deep learning technologies, human action recognition has shown significant potential in fields such as healthcare, surveillance, and human-computer interaction. However, traditional vision-based recognition techniques primarily rely on visible light or infrared cameras, making them susceptible to environmental lighting conditions and raising privacy concerns. To address these issues, this study proposes a hand movement recognition model based on ultrasonic signals and deep learning. By leveraging the reflection characteristics of sound waves, the model identifies hand movement without being affected by lighting and enhances privacy protection. This study utilizes USB ultrasonic recording devices to capture variations in sound waves generated by human movements. The signals are transformed into spectrograms using Short-Time Fourier Transform (STFT) to extract time-frequency features. A convolutional neural network (CNN) based on ResNet-50 is employed for gesture classification. A dataset consisting of five hand movement and one static background class was constructed, and the model's performance was evaluated through tests involving multiple participants. Experimental results demonstrate that the proposed model achieves a recognition accuracy of up to 95% in controlled environments and maintains a performance of 92% in cross-subject inference tests.深度學習超聲波人體動作辨識卷積神經網路短時距傅立葉轉換deep learningultrasonichuman action recognitionconvolutional neural network (CNN)short-time Fourier transform (STFT)學術論文