Improving the Performance of Micro-Doppler Spectrograms Using Deep Learning
- 발행기관 서강대학교 일반대학원
- 지도교수 Youngwook Kim
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 전자공학과
- 실제URI http://www.dcollection.net/handler/sogang/000000082459
- UCI I804:11029-000000082459
- 본문언어 영어
- 저작권 논문은 저작권에 의해 보호받습니다.
초록(요약문)
This dissertation presents a comprehensive study on enhancing the performance of intelligent sensing systems based on micro-Doppler spectrograms measured by radar sensors. While spectrograms provide rich information for applications like Human Activity Recognition (HAR) and non-contact medical monitoring, conventional analysis methodologies face fundamental challenges related to representation quality. This research proposes deep learning-based frameworks to address this issue. The impact of spectrogram representation quality on system performance is analyzed by applying a Deep Convolutional Neural Network (DCNN) to high-resolution time- frequency analysis techniques. To overcome the resolution limitations of the Short-Time Fourier Transform (STFT), the Wigner-Ville Distribution (WVD) was utilized. Applying this to the estimation of Cardiopulmonary Resuscitation (CPR) chest compression depth, it was demonstrated that the high-resolution WVD-based DCNN achieved a lower RMSE (4.47 mm) than the STFT-based DCNN (5.05 mm), proving that high-resolution spectrograms are highly effective for the precise estimation of complex signals. To address the high computational complexity of high-resolution methods like the Smoothed Pseudo WVD (SPWVD), a deep learning-based spectrogram translation model is proposed. A ResU-Net architecture is trained to perform image-to-image translation, converting computationally efficient STFT spectrograms into high-resolution SPWVD-like representations. These transformed spectrograms improved HAR classification accuracy from 80.1% (STFT) to 83.9% (transformed SPWVD). Furthermore, this approach achieved a minimal processing time of 6.5 ms, a significant reduction compared to the 140 ms required for direct SPWVD computation, thereby achieving both high performance and computational efficiency. In conclusion, this dissertation demonstrates that high-resolution spectrogram analysis is critical for sensing performance and presents a deep learning-based translation framework to achieve both high accuracy and computational efficiency. This work establishes a foundation for practical, high-performance radar sensing systems. Keywords: Micro-Doppler Spectrogram, Time-Frequency Analysis, Deep Learning, Convolutional Neural Network (CNN), ResU-Net, Wigner-Ville Distribution (WVD), Human Activity Recognition (HAR), Cardiopulmonary Resuscitation (CPR), Radar Sensor, Spectrogram Translation
more목차
Contents i
List of Tables iii
List of Figures iv
Abstract vi
1 Estimation of Compression Depth During CPR Using FMCW Radar with Deep
Convolutional Neural Network 1
1.1 Introduction 1
1.2 Related Theories 5
1.2.1 FMCW Radar Singal Model 5
1.2.2 Micro-Doppler Spectrogram 8
1.3 CPR Measurement Campaigns 11
1.4 Depth Estimation Through Doppler Integration 15
1.5 Depth Estimation Using Maximum Doppler Frequency Through Polynomial
Regression 19
1.6 Depth Estimation Using DCNN with STFT and WVD Spectrogram 23
1.6.1 DCNN to Spectrogram 23
v
1.6.2 DCNN Result Using STFT 25
1.6.3 DCNN Result Using WVD 26
1.7 Conclusion 29
1.8 Discussion 30
2 Improving Classification Accuracy of Human Micro-Doppler Signatures Measured
by Radar Sensor Using STFT to SPWVD Transformation With U-Net 32
2.1 Introduction 32
2.2 Training Data Acquisition Through Radar Simulation 35
2.2.1 Simulation-Based Training Data Generation 35
2.2.2 ResU-Net Architecture for Spectrogram Translation 38
2.3 Verification Using Measured Data 41
2.4 Classification Using CNN 42
2.5 Conclusion 44
Bibliography 46
Abstract (In Korean) 53
vi
