Skip to main content
Back

Constructor University's Fourier SOFT Tracks Moving Cars Through Rain Without Cameras

A radar processing method that works in the frequency domain promises robust odometry and simultaneous multi-object tracking for high-speed autonomous driving.

  • New paper proposes processing radar in the frequency domain rather than extracting features for odometry.
  • Method called Fourier SOFT in 2D (FS2D) decouples rotation and translation using spherical harmonics and SOFT.
  • Correlation output exposes ego motion plus every other moving object as separate peaks.
  • Tested on the Boreas dataset for radar-only odometry with no sensor fusion.
  • Motivating use case is autonomous racing overtaking maneuvers with high scene dynamics.
  • Currently 2D only and positioned as initial results, not a finished production system.

Radar has always been the sensor of last resort for autonomous vehicles operating in rain, snow, or dust, where cameras and lidar struggle. Raw radar data is noisy and sparse, and the typical fix is to extract discrete features that a matching algorithm can chew on. A new paper from a team at Constructor University argues that this feature extraction step is exactly where things go wrong, and proposes handling radar entirely in the frequency domain instead.

The work, Towards Multi-Object-Tracking with Radar, introduces initial results for a technique called Fourier SOFT in 2D (FS2D) applied to automotive radar. The authors demonstrate radar-only odometry on the Boreas dataset, which includes over 350km of driving data recorded with a 360 degree Navtech CIR304-H scanning radar across seasons and adverse weather like rain and falling snow. Their pitch is that a single mathematical pipeline can simultaneously estimate the vehicle's own motion and pick out every other moving object in the scene.

Why features fall apart at speed

Most existing radar odometry systems detect keypoints in a scan, describe them, and match them across frames. That works when the scene is mostly static and the sensor moves slowly. It falls apart when you have what the paper calls high dynamics: the car is moving fast, other vehicles are moving fast in different directions, and the returns are already smeared by noise and multipath reflections. Feature detectors start hallucinating or missing points, and downstream tracking follows them off a cliff.