Post by Harun SEVİNÇ, M. Sc.

I am very happy to share that our latest research, "Comparative Analysis of Time-Frequency Transformations and Deep Learning Architectures in the Classification of FMCW Radar Micro-Doppler Signatures," has just been published. In this study, we explored Human Activity Recognition (HAR) using mmWave FMCW radar. We specifically investigated how different time-frequency representations (STFT, CWT, and WVD) influence the spatial perception of deep learning filters under limited data constraints. To evaluate this, we designed custom lightweight architectures aimed at maximizing feature extraction efficiency without overfitting. Key Observations from Our Study: The logarithmic scaling of CWT may actually introduce unexpected challenges for standard CNN filters due to spatial warping. Standard STFT continues to offer a solid and reliable baseline for these tasks. Our CNN-Wide architecture, when combined with the Wigner-Ville Distribution, showed a strong ability to process high-resolution cross-term textures, yielding the most balanced and highest performance in our experiments. Notably, the WVD-based model identified falling activities with high precision, highlighting its promising potential for autonomous elderly care and health monitoring applications. A huge thanks to my academic advisor Prof. Dr. Levent Seyfi and the reviewers for their invaluable feedback throughout this journey. You can read the full paper via the link below. https://lnkd.in/dveDJGtc #FMCWRadar #DeepLearning #SignalProcessing #HumanActivityRecognition #MicroDoppler #MachineLearning #ElderlyCare #Research