PhD - Leuven | More than two weeks ago
Frequency Modulated Continuous Wave (FMCW) Multiple Input Multiple Output (MIMO) radar imaging is increasingly becoming a critical component in the automotive industry . As in any computational imaging system, we need to solve an image reconstruction problem after mathematically relating the unknown reflectivity distribution of the scene to the radar measurements. This requires solving an inverse problem to estimate the unknown 3D reflectivity image from the noisy radar measurements, for which both analytical and deep learning models have been proposed. The conceptually simplest reconstruction methods are the matched filter (MF) and the well-known back-projection (BP) , having high (MF) or medium (BP) complexity. However, the MF, BP and their simpler 2D/3D FFT-based approximations are not robust to noise as they directly aim at solving the equation that models the observations without considering the effect of noise or exploiting prior information. On the other hand, regularized iterative reconstruction methods incorporate additional prior knowledge (such as sparsity prior) into the reconstruction process. Total variation-based reconstruction can be given as an example of this class of methods. Such methods provide better reconstruction quality but with higher computational cost.
that exploit deep learning have emerged as an alternative to analytical methods
[1, 3]. These methods are shown to simultaneously achieve high reconstruction
quality and low computational cost for various inverse problems in imaging. The
existing deep learning-based approaches in the literature can be grouped into
three main classes: 1) learning-based direct inversion, 2) plug-and-play regularization
[4, 5], and 3) learned iterative reconstruction based on unrolling [6,7, 8].
However, deep learning-based reconstruction methods have not been studied much
in the literature in the context of forward-looking MIMO radar imaging.
This thesis aims to develop
deep learning-based fast reconstruction methods for forward-looking MIMO radar
imaging and investigate the respective merits and drawbacks of these developed
methods in comparison to the existing methods we developed at IMEC .
 Gao, X.; Roy, S.; Xing, G. MIMO-SAR: A Hierarchical High-Resolution Imaging Algorithm for mmWave FMCW Radar in Autonomous Driving. IEEE Trans. Veh. Technol. 2021, 70, 7322–7334
 4. A. Albaba, M. Bauduin, T. Verbelen, H. Sahli and A. Bourdoux, "Forward-Looking MIMO-SAR for Enhanced Radar Imaging in Autonomous Mobile Robots”, IEEE Access, 2023, doi: 10.1109/ACCESS.2023.3291611
 E. Mason, B. Yonel and B. Yazici, "Deep learning for radar," 2017 IEEE Radar Conference (RadarConf), Seattle, WA, USA, 2017, pp. 1703-1708, doi: 10.1109/RADAR.2017.7944481.
 Y. -J. Li, S. Hunt, J. Park, M. O'Toole and K. Kitani, "Azimuth Super-Resolution for FMCW Radar in Autonomous Driving," 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 17504-17513, doi: 10.1109/CVPR52729.2023.01679.
 Y. Wang, Z. He, X. Zhan, Q. Zeng and Y. Hu, "A 3-D Sparse SAR Imaging Method Based on Plug-and-Play," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, 2022, pp. 1-14, Art no. 5238514, doi: 10.1109/TGRS.2022.3221934.
 Riccardo Mazzieri and Jacopo Pegoraro and Michele Rossi , Enhanced Attention-Based Unrolling for Sparse Sequential micro-Doppler Reconstruction, 2023, arXiv, doi:10.48550/arXiv.2306.14233
 Kotte, V. V., Gishkori, S., Masood, M., & Al-Naffouri, T.Y. (2022). Enhanced Imaging for Forward Looking MIMO SAR Via Un-Supervised Deep Basis Pursuit. 2022 IEEE Radar Conference (RadarConf22). https://doi.org/10.1109/radarconf2248738.2022.9764257
 H. Mansour, S. Lohit and P. T. Boufounos, "Distributed Radar Autofocus Imaging Using Deep Priors," 2022 IEEE International Conference on Image Processing (ICIP), Bordeaux, France, 2022, pp. 2511-2515, doi: 10.1109/ICIP46576.2022.9897332.
Required background: Signal processing for wireless comm or radar. Good Knowledge in Machine Learning, Deep Learning and mathematics behind inverse problems and imaging. Fluency in Matlab and Python and some experience with ML frameworks such as Pytorch, Keras, Tensorflow
Type of work: 20% literature/theory, 60% Design and training deep learning models, Design and validate evaluation metrics for the developed models modelling/simulation, 20% Design and execute protocols for collecting data.
Supervisor: Hichem Sahli
Daily advisor: Hamed Javadi, Marc Bauduin
The reference code for this position is 2024-065. Mention this reference code on your application form.