An Integrated Framework for Human Sensing Using Deep Learning with a FMCW Radar : Presence Detection, Activity Classification, and Frame Data Reconstruction.
FMCW 레이다와 딥러닝을 활용한 인체 감지 통합 프레임워크: 존재 감지, 활동 분류 및 프레임 데이터 복원
- 주제(키워드) FMCW radar , deep learning , human sensing , human presence detection , human activity recognition , data reconstruction , feature fusion , LSTM-2D-CNN , 3D U-Net , range-Doppler map
- 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/000000082267
- UCI I804:11029-000000082267
- 본문언어 영어
- 저작권 논문은 저작권에 의해 보호받습니다.
초록(요약문)
This paper proposes an integrated deep learning-based framework for human sensing using a Frequency-Modulated Continuous-Wave (FMCW) radar embedded in a mobile device, targeting close-proximity user environments. Radar sensors are emerging as a robust alternative to camera-based systems for human sensing applications, offering key advantages such as privacy preservation and robust performance regardless of lighting or weather conditions. This research specifically focuses on two challenging scenarios: ultra- short-range detection, and an “in-pocket” scenario, where the user carries the device. The framework comprises three core components: First, it proposes a 'Human Presence Detection' method, critical for managing electromagnetic safety compliance in mobile devices. To detect the micro-vibrations inherent to the human body, I introduce two 2D features—the phasor scatter plot and the spectrogram—in addition to conventional 1D features (magnitude/phase variance). A DCNN (Deep Convolutional Neural Network)-based fusion model combining these 2D features achieved a mean accuracy of 99.83%, significantly outperforming conventional classifiers and demonstrating the feasibility of ultra-short-range human detection. Second, the research addresses 'Human Activity Recognition (HAR)' for an "in-pocket" scenario, classifying five common activities (sitting, standing, walking, running, stair climbing). Unlike conventional studies with a fixed sensor, this setup involves a sensor that moves with the user. Using Range-Doppler (RD) map sequences as input, the performance of a 3D-CNN was compared against a hybrid LSTM-2D-CNN model. Leave-One-Subject- Out (LOSO) cross-validation revealed that the LSTM-2D-CNN generalized spatiotemporal features more effectively, achieving a test accuracy of 88.34%—a significant improvement over the 3D-CNN (55.54%). Third, the framework tackles the practical challenge of 'Frame Data Reconstruction' caused by data loss, which can occur from interference with other communication modules in a mobile device. Such data loss severely degrades the performance of dynamic activity recognition. As a solution, a 3D U-Net autoencoder is proposed to reconstruct missing frames by learning spatiotemporal context. Experiments showed that for highly dynamic activities like 'running' under a 70% data loss rate, the proposed 3D U-Net achieved a classification accuracy approximately 15% higher than a simple frame-copy method, greatly enhancing system robustness. This study successfully establishes a practical framework that integrates three essential elements for mobile device-based human sensing: presence detection, activity recognition, and data robustness, all achieved by combining FMCW radar with deep learning. Key words: FMCW radar, deep learning, human sensing, human presence detection, human activity recognition, data reconstruction, feature fusion, LSTM-2D-CNN, 3D U- Net, range-Doppler map.
more목차
Contents i
List of Tables iv
List of Figures vi
Abstract viii
1 Fundamentals of FMCW Radar 1
1.1 Principles of FMCW Radar 1
1.2 Application of FMCW Radar in Human Sensing 3
2 Human Presence Detection 5
2.1 Introduction 5
2.2 Measurement Setup and Data Acquisition 8
2.2.1 Measurements of Human Body using FMCW Radar 8
2.2.2 Measurements in Close Proximity using FMCW Radar 12
2.3 Data Processing 14
2.3.1 Magnitude Variance 14
2.3.2 Phase Variance 16
2.3.3 Phasor Scatter Plot 17
2.3.4 Spectrogram 19
2.4 Classification Methods 21
2.4.1 Conventional Binary Classifier 22
2.4.2 Proposed Method: DCNN-Based Feature Fusion Model 24
2.5 Conclusion 28
3 Human Activity Recognition 29
3.1 Introduction 29
3.2 Measurement Setup and Data Acquisition 31
3.3 Data Processing 33
3.4 Classification Methods 37
3.4.1 Classification using 3D-CNN 37
3.4.2 Classification using LSTM-2D-CNN 40
3.5 Conclusion 43
4 Frame Data Reconstruction 44
4.1 Introduction 44
4.2 Simulation Scenarios for Data Loss 45
4.3 Reconstruction Methods 47
4.3.1 Reconstruction using Previous Frame (Set B) 47
4.3.2 Proposed Method: 3D-U-Net (Set C) 48
4.3.3 Performance Evaluation 49
4.4 Conclusion 51
Bibliography 52
Abstract (In Korean) 59
