Example Strange Seed: Terrain Texture as an Implicit Velocity Estimator

Date captured: 2026-02-20

The Idea

Ground vehicles traverse textured surfaces constantly. Optical-flow vectors over terrain have a direct geometric relationship to vehicle velocity and heading. The hypothesis is that a lightweight CNN trained on flow-field statistics — rather than point-tracked features — could produce a velocity estimate that is inherently drift-free, since each frame is independent of the last.

This differs from standard visual odometry: instead of tracking keypoints across frames, the model learns the statistical texture signature of motion (blur direction, spatial frequency shift, motion-field divergence) and directly regresses velocity from a single flow field.

Why It Might Work

  • Terrain texture carries dense motion information that is largely ignored by sparse feature trackers
  • Flow-field statistics are theoretically drift-free (no accumulated integration error)
  • CNNs have demonstrated surprising ability to extract subtle spatial signals from image statistics
  • Ground vehicles operate in a constrained motion manifold (mostly 2D), reducing the regression complexity

Why It Might Fail

  • Textureless or repetitive terrain (wet pavement, snow, sand) would likely break the signal entirely
  • Illumination changes and shadows introduce confounds that are hard to disentangle from velocity signals
  • The network would require a large, diverse training set paired with ground-truth velocity — expensive to collect
  • High speeds introduce motion blur that may corrupt the very statistics the model relies on
  • Not obviously better than existing methods (LiDAR odometry, IMU + wheel encoder fusion) in normal conditions

Cheap Next Test

  1. Collect a small dataset: Drive a ground vehicle at controlled speeds (0.5–3 m/s) over varied terrain with a downward-facing camera and a high-accuracy GPS/IMU reference.
  2. Compute optical flow on each frame pair using a standard algorithm (RAFT or Farnebäck).
  3. Extract flow statistics (mean magnitude, dominant direction, spatial frequency histogram) as a feature vector.
  4. Fit a linear regressor from flow statistics to ground-truth speed.
  5. Evaluate correlation: If R² > 0.8 on held-out terrain types, the signal is real and worth pursuing with a proper network.

Total cost: a few hours of data collection + an afternoon of Python. No new hardware required.

Uncertainty Level

High — this is an untested hypothesis. No literature search has been done to confirm or refute it. Treat all claims above as speculation until the cheap next test is run.

Tags: strange, state-estimation, perception