When Does Geometry Suffice? : Single-Frame Tangential Velocity Recovery from Multi-Sensor FMCW LiDAR

Sammanfattning

Frequency-Modulated Continuous-Wave (FMCW) Light Detection and Ranging (LiDAR) sensors measure per-point radial velocity through the Doppler effect, but tangential velocity (the component perpendicular to the sensor beam) is unobservable to a single sensor under classical geometry. Road vehicles operate predominantly in a horizontal plane, so velocity recovery is restricted here to the planar components (vx, vy); the vertical component vz is assumed negligible. With a multi-sensor rig and sufficient per-object line-of-sight angular diversity, tangential recovery becomes a well-posed linear problem; in the single-sensor case, learned priors over local geometry, semantic class, and beam configuration are the main alternative. Multi-frame scene-flow methods take a third route by accumulating temporal evidence across frames, but they introduce at least 100 ms of latency at 10 Hz and cannot estimate velocity before a second frame is observed. This thesis evaluates these three routes on the AevaScenes v0.1 six-sensor FMCW LiDAR dataset. The main contribution is an oracle-box multi-sensor rigid-body Doppler solver that uses ground-truth object boxes for point grouping and conditions the solve on per-object angular diversity. It achieves tangential endpoint error of 1.61 m/s on the held-out test split, within 3% of fine-tuned DeltaFlow (1.57 m/s), without requiring a prior frame. A single-sensor learned Point Transformer V3 model reduces validation tangential endpoint error from 3.14 to 2.15 m/s, showing that spatial context adds value over Doppler-only geometry. The six-sensor learned variant improves full-vector dynamic error but does not surpass the oracle-box geometric solver. These results indicate that classical multi-sensor geometry is the strongest option when object grouping and angular diversity are available, while learned priors are most useful in the single-sensor regime where tangential velocity is not geometrically observable.