Congratulations to our #ESR 9 from #MENELAOSNT, Dr. Saravanan Nagesh, who successfully defended his #PhD #thesis on the topic of “Coded waveforms for colocated MIMO radar using sparse modelling” on March 27, 2025 at the Fraunhofer-Gesellschaft – Universität Siegen, Germany under the supervision of Prof. Ender Joachim and Dr. Miguel Heredia Conde .
Topic Summary:
Advanced Driver Assistance Systems (ADAS) for autonomous driving depend on precise environmental sensing
to enable safe navigation, prevent collisions, and support reliable decision making. In real world scenarios, no
single sensor can overcome all challenges. To address this, modern ADAS systems incorporate a combination of
sensors, including cameras, LiDAR, ultrasound, and radar, each contributing unique strengths. Among these,
radar is particularly valued for its robustness in adverse weather conditions, making it indispensable for reliable
system performance. Specifically, collocated Multiple Input Multiple Output (MIMO) radar systems are the most
suitable choice for ADAS applications, offering a compact size, cost efficiency, and the ability to simultaneously
provide spatial information, range, and velocity estimation.
However, conventional automotive radars face significant challenges stemming from low resolution, which
complicates the detection of closely spaced objects or targets. These limitations hinder the fusion of multi
sensor data for reliable ADAS functionality and restrict the development of fully autonomous capabilities.
Traditional radar designs rely on uniformly spaced array configurations, which inherently tie spatial resolution
to aperture size and the number of physical antenna elements. Additionally, existing waveform candidates
often suffer from interference or elevated sidelobes, leading to scenarios with masking effects. Efforts to achieve
higher resolution without sidelobe artifacts typically result in increased costs, greater data storage requirements,
and higher computational complexity, reducing their practicality in real world applications.
One notable aspect of MIMO radars is that the measured scene is inherently sparse. Sparsity refers to a property
of signals or datasets where only a small number of elements (or features) carry significant information, while
the rest are zero or near zero. This characteristic is fundamental to advanced signal processing methods like
compressed sensing (CS), which enables high dimensional data to be represented and recovered using far
fewer measurements than traditionally required. While most studies emphasize exploiting sparsity in the scene
domain, focusing on reconstructing relevant features in 2D MIMO scenes such as Range-Angle (RA) or Doppler-
Angle (DA), sparsity in the signal domain often remains under explored. Signal domain sparsity leverages
inherent redundancies in raw radar signals, enabling more efficient data acquisition and reconstruction. By
reducing these redundancies in signal acquisition, CS methods can be applied more effectively. This approach
opens the door to replacing conventional reconstruction blocks with CS based architectures, enabling the
development of efficient CS-MIMO radars that optimize both sensing and computational efficiency.
CS holds great promise for MIMO radar systems, enabling high resolution sensing with fewer measurements.
However, its practical implementation faces key challenges, including reconstruction errors and computational
load. Central to these issues is the sensing matrix, which links the radar system to the scene. High resolution
requirements often necessitate large sensing matrices, increasing storage demands and computational com-
plexity. Additionally, gridding errors caused by mismatches between the assumed model and real world scenes
lead to reconstruction artifacts, false alarms, and reduced efficiency, further limiting reliability.
Most studies address these challenges by designing algorithms that incorporate random projections into exist-
ing system models, providing incremental improvements in CS based reconstruction. While these methods
enhance certain aspects of performance, they often rely on adapting conventional system designs rather than
fundamentally rethinking the system architecture. In this study, we take a different approach by exploring
a strategy that optimizes the radar system itself for improved CS performance. This involves replacing con-
ventional signal processing frameworks with CS specific sensing matrices, designed to minimize gridding
errors and computational demands. By treating system level parameters such as array geometry and waveform
configurations as variables for optimization, we aim to create a radar architecture inherently suited for CS based
processing, enhancing both reconstruction accuracy and system efficiency.
We propose a mutual coherence minimization approach by treating waveform and array configurations as key
parameters for optimization. Two algorithms are introduced: one to independently optimize the modulating
phase components of the coding sequences of the waveform configuration, and another to optimize the
positional placement of array elements. Unlike conventional studies, which rely on the Frobenius norm
∥G −I∥F of the Gramian G of the sensing matrix and random sub sampling, we adopt a more aggressive strategy
using the infinity norm |G −I∥∞. This approach, combined with a Temperature Guided Perturbation method
termed Annealed Projections, effectively optimizes both waveform and array configurations. The resultant waveforms exhibit superior auto and cross correlation properties, achieving lower sidelobe levels compared
to state of the art techniques. Similarly, the optimized array offers enhanced resolution and reconstruction
performance, particularly within the CS formulation.
Building on this, we further improve computational efficiency with a novel strategy called Sparse Annealed
Projections (SAP). This weighted optimization technique imposes signal sparsity by exploiting redundancies in
the system configuration. Using a priori information, SAP identifies the optimal number of spatial samples,
reducing storage requirements associated with the sensing matrix size. This proposed methodology ensures
seamless incorporation of CS into MIMO architectures, achieving reduced computational loads, compact
storage requirements, and improved overall performance.
To validate the proposed methodology, we conducted experiments on synthetic data using Monte Carlo simula-
tions. The results demonstrate significant improvements in scene reconstruction performance, providing a
proof of concept for designing sub Nyquist CDMA MIMO radar architectures for automotive imaging radar sys-
tems. This study establishes a robust framework for designing radar systems that balance high performance with
computational efficiency, paving the way for advancements in ADAS and autonomous driving technologies.
Recent Comments