Why a 15° Radar Can Produce Fine-Grained Point Clouds: Treating Artifacts as Information

Apr 16, 2026

9 min read

6 views

85%

Are we measuring the world, or are we learning to read the traces our instruments leave behind? What if the blur that makes a compact radar look crude is actually a rich code of geometry and motion, waiting for the right interpreter?

Cameras and lidars routinely boast angular resolution on the order of 0.01 degrees and 0.1 degrees respectively. Compact millimeter wave radars commonly report angular resolution near 15 degrees. On the face of it, that gap is a gulf. Yet engineers and researchers are now producing surprisingly detailed point clouds from small radars. The secret is not a miracle hardware improvement. It is a shift in perspective: from treating smear and sidelobes as errors to treating them as structured information, and then designing processing pipelines that extract that information rather than erasing it.

This article argues that high fidelity perception from compact radar is not a hardware miracle but a software reframe. I will show why artifacts encode geometry, propose a simple mental model for how to read them, and give a concrete, actionable pipeline you can apply to turn noisy range azimuth heatmaps into useful point clouds.

Setup: The tension between physics and perception

Compact millimeter wave radars are attractive: they penetrate smoke, work in darkness, see through some occlusions, and are cheap and power efficient. Those advantages make them ideal for many embedded perception tasks. Yet the physical constraints of a small aperture impose poor angular resolution. When you look at a single frame of a range azimuth heatmap from such a radar, you typically see wide blobs, smeared returns, speckle, and sidelobes. That is why radars are often dismissed as low fidelity compared with optical sensors.

The tension is straightforward: physics gives you a coarse, ambiguous measurement; applications often demand fine, localized detections. Traditional thinking treats the radar measurement as a degraded version of the truth: apply denoising and peak picking, accept coarse angular precision, or throw hardware at the problem. That is the setup.

A different posture is possible. Instead of asking how to remove everything that looks wrong, ask what structures in the heatmap are reproducible, physically meaningful, and recoverable. That change in question opens a path toward extracting geometry that seems to outrun the advertised angular resolution.

Exploration: Why artifacts can betray truth

Three properties of compact radar measurements make them both challenging and useful.

Measurements are low resolution in angle but high fidelity in range. Range bins are often fine enough to locate returns precisely in distance. Angular smear compresses lateral structure into patterns across azimuth bins, but those patterns are systematic and repeatable.
Physics creates structured artifacts. Sidelobes, speckle, and multipath are not random dust. They are the result of aperture geometry, waveform, and scene interaction. When a point reflector is measured by a finite aperture, the point spread function modulates the true scene. If you know or can learn that modulation, you can invert it.
Motion and time are amplifiers. Multiple frames captured as the sensor or targets move effectively synthesize a larger aperture. Temporal coherence and motion parallax convert ambiguous 2D blobs into discriminable trajectories.

Think of the heatmap as an impressionist painting rather than a failed photograph. Each brush stroke is messy, but the strokes follow underlying geometry. A viewer who knows how brush strokes relate to the scene can reconstruct sharp edges that the strokes themselves never directly show.

Analogy: imagine listening to a crowded room through a muffled microphone. A naive denoiser might remove all the ambient grain and leave you with a few loud voices. A better approach models how the microphone colors sound and uses overlaps in frequency and time to separate speakers. The muffledness is not just noise; it is a convolution of source signals with a device response. In radar, sidelobes and smear are the radar convolutional signature.

The blur you see in a range azimuth image is a coded imprint of the world, not the world itself. If you learn the code, you can read details that the hardware resolution seems to forbid.

This reframing has consequences for processing design. In particular, it implies three guiding principles:

Preserve structured artifacts early in the pipeline. Aggressive smoothing or conservative thresholding that collapses speckle removes information you may need.
Move from thresholding to selective thresholding. You want to suppress sensor noise while retaining side information that correlates with real geometry.
Combine spatial models with temporal or motion cues. Single frames are ambiguous; sequences can disambiguate.

Synthesis: A practical framework for reading heatmaps

Below is a compact pipeline you can apply to transform raw range azimuth heatmaps into higher resolution point clouds or detections. Each step keeps the central insight in view: treat artifacts as recoverable data when possible.

Acquire a clean, calibrated range azimuth heatmap
- Use consistent preprocessing so that bin magnitudes are comparable across frames. That means calibrating gain and compensating for known antenna phase offsets.
- Visualize heatmaps early while keeping raw data accessible. Many community tools plot heatmaps for debugging; adopt similar visual checks to build intuition about how returns appear for canonical targets.
Do low level adaptive thresholding, not blunt denoising
- Apply a low thresholding rule that removes outlier noise but leaves speckle patterns intact. Think of this as denoising that preserves texture rather than smoothing it away.
- Avoid large morphological closing or heavy Gaussian blurs at this stage. Those operations remove the very structure you want to interpret.
Extract local structures via anisotropic filters and connected components
- Use filters that follow the orientation of streaks and lobes in the heatmap. Speckle often has anisotropic shapes because of the underlying aperture. Anisotropic diffusion or oriented morphological filters can separate structure from random noise.
- Identify connected clusters across range and azimuth. Even when a target projects as a wide azimuth blob, its cluster shape and intensity profile carry cues about lateral position and extent.
Build a physical forward model and perform sparse inversion when possible
- Formalize the radar point spread function for your aperture. Even a coarse PSF yields a convolution model: measured heatmap equals PSF convolved with scene reflectivity plus noise.
- Use sparse recovery or deconvolution techniques to estimate reflectivity. Compressed sensing approaches perform well when the scene is sparse in an appropriate basis. Regularize toward sparsity in angular domain to recover localized reflectors.
Fuse frames with motion compensation to synthesize aperture
- Track ego motion or estimate it from inertial sensors. Compensate for sensor movement and integrate multiple frames to build an effective longer aperture.
- For moving targets, use short time window alignment via estimated velocity to stack returns coherently. Motion parallax resolves lateral ambiguities.
Train discriminative models on heatmaps, not on denoised peaks
- Treat the thresholded heatmap as an image input to convolutional networks trained to output point clouds, bounding boxes, or semantic labels.
- Include artifact variations in training data so models learn to interpret sidelobes and smear as cues rather than nuisances. Augment training with synthetic forward model outputs to increase coverage.
Iterate with calibration and canonical tests
- Measure responses using canonical reflectors placed at known positions and distances. Use these calibration runs to refine PSF estimates and to validate inversion routines.

Concrete example: suppose you have a radar with 15 degree nominal angular resolution. A single static pedestrian may appear as a broad azimuth smear without a sharp peak. After low level thresholding and anisotropic filtering, the cluster of bins belonging to that pedestrian will show a consistent range profile and a characteristic intensity gradient. Feeding a short sequence of such heatmaps into a sparse recovery routine, aligned using small ego motion, can make the lateral estimate much narrower than 15 degrees. Adding a neural model trained to expect the PSF shaped smear can further sharpen the lateral localization into a tight point cloud representation.

Why does this work? Because angular resolution quoted for hardware is a single number defined under ideal assumptions. Real scenes are not worst case. Range fidelity, scene sparsity, motion, and structured artifacts combine so that the effective localization achievable with intelligent processing can be substantially finer than the raw angular resolution suggests.

Actionable insights: what to try tomorrow

Here are practical steps you can apply now to start extracting more from compact radar heatmaps.

Start with low level thresholding: set a threshold that is permissive enough to keep speckle patterns. Do not apply aggressive smoothing before analysis.
Treat the heatmap as an image: use convolutional filters and image processing tools to extract oriented structures and connected regions.
Collect short sequences and perform motion compensated stacking. Even modest motion over tens of frames synthesizes a larger aperture.
Build or estimate a simple point spread function and experiment with deconvolution or sparse inversion. Regularize for sparsity in angle.
Train a learning based model directly on thresholded heatmaps to map to point clouds or detections, using synthetic PSF augmented data if real labeled data is scarce.
Validate with canonical targets at known positions to verify that your processing recovers lateral positions better than naive peak picking.

Key Takeaways

Preserve artifact structure early: low level thresholding that filters out random noise while keeping speckle and sidelobes gives you more information to work with.
Treat range azimuth heatmaps as images: image processing and convolutional models can learn to interpret smear patterns as geometric cues.
Use temporal integration with motion compensation: multiple frames can synthesize a larger aperture and disambiguate lateral position.
Model the sensor: even an approximate point spread function enables deconvolution and sparse recovery that sharpen localization beyond nominal angular resolution.
Calibrate and iterate: controlled calibration runs provide the ground truth needed to refine forward models and learning targets.

Conclusion: read the blur and design differently

The conventional view treats a radar with 15 degree angular resolution as incapable of producing fine-grained point clouds. That view is correct only if you insist on interpreting radar output the same way you interpret a high resolution camera image. If instead you read the heatmap as a coded imprint of the scene, you can invert the code using a combination of selective thresholding, physical modeling, temporal fusion, and discriminative learning.

The deeper lesson applies beyond radar: measurements are always mediated by instruments that leave structured traces. Progress occurs when we stop punishing instruments for those traces and start training interpreters to read them. For compact radar, that means treating sidelobes and speckle not as enemies but as encrypted hints. The practical payoff is real: better localization, improved detection in occlusion, and affordable sensors that deliver capability far beyond what their raw numbers advertise.

If you want sharper perception from small sensors, do not demand better hardware first. Learn to translate the language of artifacts into geometry. When you succeed, the blur resolves into signal, and the apparent boundary between hardware and perception dissolves.