Millimeter-wave (mmWave) and Frequency Modulated Continuous Wave (FMCW) radar sensors are highly valued in autonomous systems because they can see through adverse weather conditions like fog, dust, and heavy rain where optical sensors fail. However, they suffer from a fundamental physical limitation: poor angular resolution.
While a radar can determine the distance (range) and speed (Doppler) of an object with extreme precision by measuring the timing and frequency shifts of returned radio waves, resolving where that object is horizontally (azimuth) or vertically (elevation) is much harder. This spatial blurring causes distinct targets in close proximity to merge into a single indistinct blob, making tasks like object tracking, boundary estimation, and semantic segmentation incredibly challenging.
To overcome this hardware bottleneck without adding bulky, expensive physical antennas, researchers are developing algorithmic techniques to achieve angular super-resolution. By leveraging deep learning, wave physics, temporal context, and cross-modal supervision, these methods effectively "hallucinate" or reconstruct high-resolution spatial details from sparse, low-resolution radar data.
The Root of the Angular Bottleneck
To understand why radar struggles with angular resolution, we have to look at the physics of antenna arrays. In a typical Multiple-Input Multiple-Output (MIMO) radar system, the angular resolution ($\theta$) is inversely proportional to the physical size of the antenna aperture ($L$) relative to the wavelength ($\lambda$):
$$\theta \approx \frac{\lambda}{L}$$
Because compact sensors designed for vehicles or small robots must remain physically small, their aperture size $L$ is strictly constrained. As noted by Junfeng Guan et al. in Through Fog High-Resolution Imaging Using Millimeter Wave Radar, this hardware constraint limits standard mmWave radars to an angular resolution that often exceeds 15 degrees, turning sharp environmental boundaries into blurry shapes.
When a radar's angular resolution is coarse, the energy reflected from two closely spaced objects overlaps within the same spatial bin. As Arthur Ouaknine et al. demonstrate in the CARRADA Dataset, this causes clustering and perception algorithms to suffer from identity merging or severe localization errors in crowded environments.
Hallucinating Virtual Antennas Upstream
The most direct way to improve angular resolution is to increase the number of spatial samples. If we cannot add physical antennas to the chip, can we generate "virtual" ones?
Traditional signal processing applies a Fourier transform across the physical antenna channels to calculate the angle of arrival. If we only have a few antennas, this transform suffers from severe leakage and wide mainlobes. To bypass this, Yu-Jhe Li et al. introduced a method in Azimuth Super-Resolution for FMCW Radar in Autonomous Driving that moves super-resolution "upstream" to the raw Analog-to-Digital Converter (ADC) signal level. By training a lightweight neural network to predict the phase and amplitude values of missing receiver antennas directly from the raw data, they effectively synthesize a much larger virtual antenna array. This allows the system to distinctly separate closely spaced objects before the spatial map is even generated.
[Raw ADC Data] ---> [Neural Network] ---> [Expanded Virtual Aperture] ---> [Sharp Spatial Map]
(Sparse Channels) (Hallucination) (Many Virtual Channels) (High Resolution)
Taking this physics-centric approach further, Yinyan Bu et al. proposed NEAR: Neural Electromagnetic Array Response. Instead of relying on massive offline training datasets, the authors embed the exact physics of electromagnetic wave propagation directly into an untrained Implicit Neural Representation (INR). By optimizing the network to comply with wave physics, it accurately predicts missing antenna responses from just a single snapshot, proving that physical priors can substitute for missing hardware.
Trading Time and Motion for Space
If the radar sensor is mounted on a moving platform—such as an autonomous car or a drone—we can exploit that motion to solve the spatial resolution problem. This is the core intuition behind Synthetic Aperture Radar (SAR) and Doppler Tomography.
As a radar moves, it transmits pulses and receives reflections from different physical locations over time. A target at a specific angle will exhibit a unique history of phase shifts and Doppler frequencies relative to the sensor's velocity vector. By mathematically aggregating these observations across a temporal window, the system treats the path traveled by the sensor as one massive "synthetic" antenna array.
In DART: Implicit Doppler Tomography for Radar Novel View Synthesis, John Miller et al. leverage this concept by combining Doppler shifts with Neural Radiance Fields (NeRFs). Because the Doppler frequency of a return is intrinsically tied to the target's relative angle, the network can utilize the sensor's motion to disentangle ambiguous spatial returns, generating high-fidelity 3D representations from otherwise blurry scans. Similarly, the authors of Performance Analysis of Automotive SAR With Radar Based Motion Estimation demonstrate that synthetic aperture techniques allow standard automotive radars to resolve fine spatial details—like distinguishing between curbs, poles, and parked cars—without altering the physical sensor size.
Cross-Modal Supervision: Learning from Cameras and LiDAR
Another powerful strategy for beating the angular resolution limit is to use cross-modal deep learning. During the training phase, a radar network is paired with high-resolution optical sensors (like cameras or LiDAR). The optical sensors act as a "teacher," providing sharp, pixel-level geometric boundaries that the radar lacks.
The core challenge here is dealing with the wide beamwidth of the radar. As Y. F. Long explains in Radar-Camera Pixel Depth Association for Depth Completion, a single radar return could correspond to any number of visual pixels within a broad angular neighborhood. To resolve this ambiguity, the author developed a pixel-to-radar association framework that learns to map sparse, wide radar beams to their precise visual counterparts.
Radar Beam: [================= Blurry Return =================]
|
Camera View: [ Object A ] | [ Object B ]
v
Association: +---> [Originates precisely here!]
By learning these associations, deep networks can recognize the subtle "signatures" within a blurry radar return that indicate distinct objects. For instance, in RODNet, Yizhou Wang 0005 et al. utilize a camera-radar fusion model to cross-supervise a radar-only student network. This training process teaches the radar-only network to successfully separate the reflections of small-cross-section targets, like pedestrians, even when they are physically merging with the reflections of larger nearby vehicles.
Key Foundational Work
A detecting and identifying method for two stationary human targets using single-channel ultra-wideband radar (2012) — contributed 1 concept elements to this direction.
A neural network approach to target classification for active safety system using microwave radar (2010) — contributed 1 concept elements to this direction.