簡易檢索 / 詳目顯示

回結果列表
研究生： 吳昱泓
Wu, Yu-Hong
論文名稱： 基於稀疏點雲學習與雷達對齊幾何適配之高效率 FMCW 雷達人體動作辨識
Sparse Point-Cloud Learning and Radar-Aligned Geometric Adaptation for Efficient FMCW Radar Human Activity Recognition
指導教授： 楊慶隆
Yang, Chin-Lung
學位類別： 碩士
Master
系所名稱： 電機資訊學院 - 電機工程學系
Department of Electrical Engineering
論文出版年： 2026
畢業學年度： 114
語文別： 中文
論文頁數： 100
中文關鍵詞： 調頻連續波雷達 、人體動作辨識 、稀疏點雲 、壓縮感測 、點雲深度學習 、快速迭代收縮閾值演算法 、距離–速度–時間幾何對齊
外文關鍵詞： compressed sensing, FISTA, FMCW radar, human activity recognition, point-cloud deep learning, RVT geometric alignment,, sparse point cloud
相關次數： 點閱：85  下載：2 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 本研究提出一套以稀疏性為核心的兩階段調頻連續波（Frequency Modulated Continuous Wave, FMCW）雷達人體動作辨識（Human Activity Recognition, HAR）管線，以稀疏點雲取代密集頻譜圖，在邊緣裝置上兼顧準確率、參數量與推論速度。
    第一階段以 LASSO 正則化搭配快速迭代收縮閾值演算法（Fast Iterative Shrinkage-Thresholding Algorithm, FISTA）進行稀疏重建，採時間–都卜勒（Time-Doppler, TD）域 Gabor 字典與距離–都卜勒（Range-Doppler, RD）域離散傅立葉變換（Discrete Fourier Transform, DFT）字典，生成時間域稀疏點（Time-Domain Sparse Point, TDSP）與動態距離–都卜勒稀疏點（Dynamic Range-Doppler Sparse Point, DRDSP）點雲。本研究比較 OMP、CoSaMP、ISTA 與 FISTA 在 5.8 GHz、24 GHz 與 77 GHz 三種字典規模下的精度–速度折衷，確認 FISTA 在毫米波大字典場景下具最佳跨頻段可擴展性。
    第二階段針對雷達距離–速度–時間（Range-Velocity-Time, RVT）空間提出兩種客製化點雲骨幹。RVT-PointMLP 以 RVT 各向異性度量修正 k 近鄰搜尋，並搭配幾何仿射（Geometric Affine, GA）模組進行可學習局部正規化。RVT-PointODE 以神經常微分方程（Neural Ordinary Differential Equation, Neural ODE）共享 block 取代多層殘差堆疊，降低浮點運算量（Floating Point Operations, FLOPs），並以振幅加權統計、條件去相關剪切與分塊逐點正規化之 RVT-aware 局部分組模組強化幾何建模能力。
    實驗以分層五折交叉驗證（60/20/20）於 CI4R 77 GHz、24 GHz 與 5.8 GHz FMCW 三公開資料集上評估。RVT-PointODE 分別達 87.12%、83.11% 與 95.76% 之整體準確率（Overall Accuracy, OA），RVT-PointMLP 為 78.86%、82.08% 與 94.86%，且此效能增益於三個頻段間穩定重現。以 5.8 GHz 為例，RVT-PointODE（152K 參數、153M FLOPs）之參數量、FLOPs 與峰值顯示卡記憶體（Video Random Access Memory, VRAM）用量分別約為 PointNet 基線的三分之一、七分之一與六分之一，驗證稀疏點雲管線於資源受限邊緣雷達系統之部署潛力。

    This thesis presents a sparsity-driven, two-stage FMCW radar Human Activity Recognition (HAR) pipeline for elderly in-home monitoring that replaces dense micro-Doppler spectrograms with compact sparse point clouds, balancing accuracy against parameter count, computational cost, and memory footprint required for edge deployment.
    Stage one recasts sparse recovery as a LASSO problem solved with the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA), using a Gabor dictionary in the Time-Doppler domain and a DFT dictionary in the Range-Doppler domain to generate Time-Domain Sparse Point (TDSP) and Dynamic Range-Doppler Sparse Point (DRDSP) clouds; a cross-band comparison against OMP, CoSaMP, and ISTA confirms FISTA offers the best accuracy-speed trade-off and scalability.
    Stage two introduces two radar-customized point-cloud backbones operating on the Range-Velocity-Time (RVT) space. RVT-PointMLP replaces the Euclidean $k$-NN metric with a learnable RVT anisotropic metric and augments PointMLP's ResP blocks with a Geometric Affine module; RVT-PointODE further replaces the residual block stack with a single shared Neural ODE block, substantially reducing parameters while preserving the same geometry correction.
    Under stratified five-fold cross-validation, RVT-PointODE attains 87.12%, 83.11%, and 95.76% overall accuracy on the CI4R 77 GHz, 24 GHz, and 5.8 GHz datasets respectively, outperforming RVT-PointMLP across all bands, while needing only about one-third the parameters, one-seventh the FLOPs, and one-sixth the peak VRAM of an in-protocol PointNet baseline on 5.8 GHz, confirming a practical route toward low-power edge deployment.

    摘要 i Extended Abstract ii 誌謝 vii 目錄 viii 表目錄 xi 圖目錄 xii 縮寫總表 xiv 第一章 緒論 1 1.1 研究背景與動機 1 1.2 相關研究回顧 3 1.2.1. FMCW 雷達與頻譜圖導向之深度學習方法 3 1.2.2. 稀疏表徵與點雲式雷達 HAR 4 1.2.3. 點雲深度學習網路 5 1.3 研究問題與目標 7 1.4 研究貢獻 9 1.5 論文架構說明 10 第二章 理論基礎 12 2.1 頻率調變連續波雷達 12 2.1.1. FMCW 雷達信號數學模型 12 2.1.2. 三維雷達資料立方體（Radar Data Cube）建構 14 2.1.3. 距離與都卜勒處理 16 2.1.4. 靜態雜波與前景運動分離 16 2.1.5. 從雷達影像到稀疏點雲 17 2.2 壓縮感測理論 18 2.2.1. 量測模型與稀疏性 18 2.2.2. 理論保證條件（RIP 與不相干性） 19 2.2.3. 稀疏重建問題 19 2.2.3.1. 貪婪演算法：OMP 與 CoSaMP 20 2.2.3.2. 凸鬆弛：基追蹤與 LASSO 23 2.2.4. LASSO 正則化與本研究設定 23 2.2.5. 迭代收縮閾值演算法（ISTA） 24 2.2.6. 快速迭代收縮閾值演算法（FISTA） 26 2.2.7. 學習式稀疏恢復：LISTA 與演算法展開 27 2.3 點雲幾何學習的基礎操作 28 2.3.1. 最遠點採樣（Farthest Point Sampling, FPS） 28 2.3.2. 𝑘 近鄰搜尋（𝑘-Nearest Neighbor Search） 28 2.3.3. 局部特徵聚合（Local Feature Aggregation） 29 2.3.4. 點雲骨幹的局部正規化設計 29 2.4 殘差連接與深層點雲網路訓練 30 2.5 Neural ODE 與連續深度網路 30 第三章 系統設計與所提方法 32 3.1 系統整體架構 32 3.2 稀疏字典設計與雷達對應 33 3.2.1. 時間域 Gabor 字典（TD） 33 3.2.2. 距離–都卜勒 DFT 字典（RD） 33 3.3 FISTA 回溯線搜索步長設計 34 3.4 稀疏 RVT 點雲生成 35 3.5 RVT-PointMLP 架構設計 36 3.5.1. 理論基礎：雷達相位空間的橢圓局部化 36 3.5.2. 雷達對齊之 RVT 度量 37 3.5.3. 幾何仿射模組（Geometric Affine, GA） 38 3.6 RVT-PointODE 架構設計 39 3.6.1. RVT-aware 局部分組 40 3.6.2. ODEPBlock 與輕量化機制 41 3.7 雷達感知資料增強策略 41 第四章 實驗設計與結果分析 43 4.1 資料集介紹 43 4.1.1. CI4R 77 GHz 雷達活動資料集 43 4.1.2. CI4R 24 GHz 雷達活動資料集 47 4.1.3. 5.8 GHz FMCW 稀疏點雲活動資料集 47 4.2 稀疏重建方法驗證 48 4.2.1. 時域稀疏字典選擇 48 4.2.2. RD 域稀疏特徵萃取驗證 50 4.2.3. 演算法速度比較 53 4.3 實驗設定 56 4.3.1. 評估指標與比較基準 56 4.3.2. 實作細節與超參數設定 58 4.4 主要實驗結果 61 4.4.1. CI4R 77 GHz 實驗結果 61 4.4.2. CI4R 24 GHz 實驗結果 63 4.4.3. 5.8 GHz 實驗結果 64 4.5 深度分析 65 4.5.1. 消融實驗 65 4.5.2. 誤差分析 66 4.5.3. 模型超參數分析：輸入點數 𝑁pts、𝑘 與嵌入維度 𝑑 69 4.6 實際量測測試 73 4.6.1. 量測系統架構 73 4.6.2. 量測環境與流程 75 4.6.3. 量測結果與討論 76 第五章 結論與未來展望 78 5.1 結論 78 5.2 未來展望 79 參考文獻 80

    [1] H. A. Ullah, S. Letchmunan, M. S. Zia, U. M. Butt, and F. H. Hassan, "Analysis of deep neural networks for human activity recognition in videos—a systematic literature review," IEEE access, vol. 9, pp. 126366–126387, 2021.

    [2] Y. Wang, S. Cang, and H. Yu, "A survey on wearable sensor modality centred human activity recognition in health care," Expert Systems with Applications, vol. 137, pp. 167–190, 2019.

    [3] O. D. Lara and M. A. Labrador, "A survey on human activity recognition using wearable sensors," IEEE communications surveys & tutorials, vol. 15, no. 3, pp. 1192–1209, 2012.

    [4] C. Li, V. M. Lubecke, O. Boric-Lubecke, and J. Lin, "A review on recent advances in Doppler radar sensors for noncontact healthcare monitoring," IEEE Transactions on microwave theory and techniques, vol. 61, no. 5, pp. 2046–2060, 2013.

    [5] J. Zhang et al., "A survey of mmWave-based human sensing: Technology, platforms and applications," IEEE Communications Surveys & Tutorials, vol. 25, no. 4, pp. 2052–2087, 2023.

    [6] I. Ullmann, R. G. Guendel, N. C. Kruse, F. Fioranelli, and A. Yarovoy, "A survey on radar-based continuous human activity recognition," IEEE Journal of Microwaves, vol. 3, no. 3, pp. 938–950, 2023.

    [7] V. N. Dang, N. C. Hoang, M. T. Le, K. Nguyen, and Q. C. Nguyen, "Deep Learning-Based Human Activity Recognition With FMCW Radar: A Review," IEEE Sensors Journal, vol. 26, no. 4, pp. 5302–5323, 2026, doi: 10.1109/JSEN.2026.3651578.

    [8] S. Z. Gurbuz, M. M. Rahman, E. Kurtoglu, T. Macks, and F. Fioranelli, "Cross-frequency training with adversarial learning for radar micro-Doppler signature classification (Rising Researcher)," in Radar Sensor Technology XXIV, 2020, pp. 58–68.

    [9] W. Li et al., "Real-time fall detection using mmwave radar," in ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 16–20.

    [10] Y. Kim, I. Alnujaim, and D. Oh, "Human activity classification based on point clouds measured by millimeter wave MIMO radar with deep recurrent neural networks," IEEE Sensors Journal, vol. 21, no. 12, pp. 13522–13529, 2021.

    [11] F. Luo et al., "RadarAttn: Efficient Radar-Based Human Activity Recognition by Integrating Visual Attention and Self-Attention," IEEE Transactions on Networking, vol. 34, pp. 1582–1595, 2025.

    [12] Y. Wu, F. Fioranelli, and C. Gao, "RadMamba: Efficient Human Activity Recognition Through a Radar-Based Micro-Doppler-Oriented Mamba State-Space Model," IEEE Transactions on Radar Systems, vol. 4, pp. 261–272, 2025.

    [13] D. L. Donoho, "Compressed sensing," IEEE Transactions on Information Theory, vol. 52, no. 4, pp. 1289–1306, 2006.

    [14] E. J. Candés and M. B. Wakin, "An introduction to compressive sampling," IEEE Signal Processing Magazine, vol. 25, no. 2, pp. 21–30, 2008.

    [15] C. Ding, L. Zhang, H. Chen, H. Hong, X. Zhu, and F. Fioranelli, "Sparsity-based human activity recognition with PointNet using a portable FMCW radar," IEEE Internet of Things Journal, vol. 10, no. 11, pp. 10024–10037, 2023.

    [16] C. Ding, H. Chen, L. Zhang, H. Hong, and X. Zhu, "Human Activity Data with a 5.8GHz FMCW Radar." [Online]. Available: https://dx.doi.org/10.21227/aze5-h339

    [17] F. Fioranelli, S. A. Shah, H. Li, A. Shrestha, S. Yang, and J. Le Kernec, "Radar signatures of human activities." [Online]. Available: https://doi.org/10.5525/gla.researchdata.848

    [18] C. R. Qi, H. Su, K. Mo, and L. J. Guibas, "PointNet: Deep learning on point sets for 3D classification and segmentation," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 652–660.

    [19] G. Qian et al., "PointNeXt: Revisiting PointNet++ with improved training and scaling strategies," Advances in Neural Information Processing Systems, vol. 35, pp. 23192–23204, 2022.

    [20] X. Ma, C. Qin, H. You, H. Ran, and Y. Fu, "Rethinking network design and local geometry in point cloud: A simple residual MLP framework," arXiv preprint arXiv:2202.07123, 2022.

    [21] J. Li, J. Wang, and T. Xu, "Pointgl: A simple global-local framework for efficient point cloud analysis," IEEE Transactions on Multimedia, vol. 26, pp. 6931–6942, 2024.

    [22] K. Sugiura, M. Yasuda, and H. Matsutani, "PointODE: Lightweight Point Cloud Learning with Neural Ordinary Differential Equations on Edge," arXiv preprint arXiv:2506.00438, 2025.

    [23] E. C. Marques, N. Maciel, L. Naviner, H. Cai, and J. Yang, "A Review of Sparse Recovery Algorithms," IEEE Access, vol. 7, pp. 1300–1322, 2019, doi: 10.1109/ACCESS.2018.2886471.

    [24] D. Needell and J. A. Tropp, "CoSaMP: Iterative signal recovery from incomplete and inaccurate samples," Applied and computational harmonic analysis, vol. 26, no. 3, pp. 301–321, 2009.

    [25] S. L. Brunton and J. N. Kutz, Data-driven science and engineering: Machine learning, dynamical systems, and control. Cambridge University Press, 2022.

    [26] K. Gregor and Y. LeCun, "Learning fast approximations of sparse coding," in Proceedings of the 27th international conference on international conference on machine learning, 2010, pp. 399–406.

    [27] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778.

    [28] R. T. Chen, Y. Rubanova, J. Bettencourt, and D. K. Duvenaud, "Neural ordinary differential equations," Advances in neural information processing systems, vol. 31, 2018.

    [29] Y.-H. Wu and C.-L. Yang, "Sparsity-Based Range–Velocity–Time PointMLP for FMCW Radar Human Activity Classification with Efficient Computation," in IEEE MTT-S Int. Microwave Symp. (IMS), Jun. 2026.

    [30] I. Daubechies, "Time-frequency localization operators: a geometric phase space approach," IEEE transactions on information theory, vol. 34, no. 4, pp. 605–612, 2002.

    [31] P. Contributors, "Pointcept: A Codebase for Point Cloud Perception Research." 2023.

    [32] I. Loshchilov and F. Hutter, "Decoupled weight decay regularization," arXiv preprint arXiv:1711.05101, 2017.

    [33] L. N. Smith and N. Topin, "Super-convergence: Very fast training of neural networks using large learning rates," in Artificial intelligence and machine learning for multi-domain operations applications, 2019, pp. 369–386.

    [34] A. Ibrahim, M. Z. Khan, M. Imran, H. Larijani, Q. H. Abbasi, and M. Usman, "RadSpecFusion: Dynamic attention weighting for multi-radar human activity recognition," Internet of Things, p. 101682, 2025.

    [35] R. Astley and L. Morris, "At-scale impact of the Net Wok: A culinarically holistic investigation of distributed dumplings," Armenian Journal of Proceedings, vol. 61, pp. 192–219, 2020.

    [36] L. Morris and R. Astley, "Net Wok++: Taking distributed dumplings to the cloud," Armenian Journal of Proceedings, vol. 65, pp. 101–118, 2022.

    下載圖示
    校外：立即公開
    QR CODE