Full Velocity Estimation from Partial Radial Observations

This direction is split into other directions: radial velocity ambiguity. Its last version stays readable and citable here.

Version 1 of 176 papers95 conceptsfrozen 2026-10-04

Imagine standing on a train platform as an express train rushes past. If you look directly at the front of the train as it approaches, you can easily tell how fast it is coming toward you. But if you try to gauge its speed while looking perfectly perpendicular to the tracks as it flashes by, your sense of its "depth" or distance change is zero.

This is the fundamental challenge of radial velocity sensing. Most of our most robust sensors—Radar, FMCW LiDAR, and even the human ear via the Doppler effect—are physically constrained to measuring only the component of motion that moves directly toward or away from the sensor. This leaves "tangential" or "transverse" motion (moving sideways) invisible to the raw signal. Recent research across robotics, telecommunications, and computer vision is converging on a shared goal: recovering the full 3D velocity vector from these incomplete, one-dimensional observations.

The Doppler Blind Spot

The core problem is geometric. As Mian Guo notes, an object moving perfectly perpendicular to a sensor beam exhibits zero radial velocity. In autonomous driving, this is a safety nightmare: a pedestrian crossing the street in front of a car might be moving at 5 mph, but to a Doppler-based sensor, they appear as a stationary object, indistinguishable from a mailbox or a parked car.

This "blind spot" isn't limited to radar. In computer vision, Antoine Manzanera describes a variation of the "aperture problem" where radial motion toward a lens produces negligible intensity variations at the pixel level. Similarly, Hanyu Zhou and Gim Hee Lee demonstrate that event-based cameras, which trigger on brightness changes, are far less sensitive to objects moving along the optical axis than those moving laterally across the frame.

Leveraging Near-Field Physics

One of the most elegant solutions to this problem emerges from the shift toward Extremely Large Aperture Arrays (ELAA) in 6G networks. Traditionally, sensors operate under a "far-field" assumption, where radio waves are treated as flat, parallel planes. In this model, you can only measure the "push" and "pull" of the wave—the radial speed.

However, Zhaolin Wang, Xidong Mu, and Yuanwei Liu have pioneered near-field velocity sensing. By recognizing that waves are actually spherical, they show that a single base station can observe how the curvature of the wavefront changes as a target moves. This allows the system to estimate both radial and transverse velocity simultaneously. As Hao Jiang explains, this "full motion status" enables predictive beamforming—the ability for a network to "aim" its signal at where a high-speed user will be, rather than where they were.

Multi-View and Cross-Modal Fusion

If a single sensor is blind to sideways motion, the most intuitive fix is to add a second sensor at a different angle. Xianghao Yu and S. H. Song propose using Intelligent Reflecting Surfaces (IRS) to act as a second "eye," providing a different Doppler perspective that resolves the ambiguity of mono-static radar.

In the realm of human sensing, Jaeho Choi and Soheil Hor developed MVDoppler-Pose, which fuses multiple radar viewpoints to reconstruct human walking poses. Because different body parts (arms, legs) move in non-radial trajectories, a single radar would lose the signal of a swinging arm moving sideways. By combining views, they create a "trajectory-agnostic" system that maintains accuracy regardless of which way the person is walking.

Another powerful approach is cross-modal "teaching." Yunfei Long and Daniel D. Morris fused radar with camera optical flow, using the camera's strength (detecting lateral motion) to fill in the radar's blind spot. Similarly, Fangqiang Ding introduced CMFlow, which uses LiDAR and cameras to supervise radar during training, essentially teaching the radar model to "infer" the missing tangential components of scene flow from the patterns it can see.

Differential Constraints and Ray Flow

For vision-centric systems, the challenge is often the inverse: cameras see sideways motion well but struggle with depth (Z-axis) velocity. Sizhuo Ma and Mohit Gupta addressed this by deriving the ray flow equation. Instead of trying to reconstruct depth first and then calculate motion, they relate 3D scene motion directly to the gradients of a 4D light field. This allows for sub-millimeter precision in tracking motion toward or away from the lens, effectively solving the "looming" ambiguity that plagues standard 2D optical flow.

Why It Works: The Path to 5D Tracking

The ultimate goal of this research direction is what Zhuoran Li and Zhen Gao call "5D tracking": estimating 3D position, radial velocity, and tangential velocity (or orientation) simultaneously.

By moving beyond simple point-tracking and incorporating the physics of wave propagation (near-field), the geometry of multiple perspectives (multi-static), and the constraints of physical motion (physics-informed learning), researchers are turning "blind" sensors into high-fidelity observers. The result is a generation of autonomous systems that don't just see where an object is, but truly understand where it is going.

Papers in this version

Showing 60 of 76 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_full_velocity_estimation_from_partial_ra_eae8a085,
  title = {Full Velocity Estimation from Partial Radial Observations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/full-velocity-estimation-from-partial-radial-observations/rdr_eae8a0858e1744f5b5db9f805ad0cd69},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_full_velocity_estimation_from_partial_ra_eae8a085_v1,
  title = {Full Velocity Estimation from Partial Radial Observations},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/full-velocity-estimation-from-partial-radial-observations/rdr_eae8a0858e1744f5b5db9f805ad0cd69/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}