Master’s thesis · Radar perception

End-to-End AI Radar Signal Processing

For my master’s thesis, I extended T-FFTRadNet with a dense radial-velocity head, enabling joint detection, occupancy, and per-cell velocity prediction from range–Doppler radar data.

T-FFTRadNet extensionDense radial velocityPyTorch · Swin Transformer

Predicted bird’s-eye-view radar velocity grid with a mean absolute error label

01 / At a glance

Motion-aware perception in one radar model.

The work added dense scene-level motion estimation to T-FFTRadNet, an existing multi-task radar perception architecture, while retaining detection and occupancy outputs.

RoleMaster’s thesis · model extension

InputRange–Doppler radar data

OutputDetection · occupancy · radial velocity

Measured changeOccupancy TPR 64.53% → 67.01%

02 / Problem

Radar senses motion directly, but dense motion fields remain difficult.

Doppler makes radar inherently sensitive to motion, yet many learned pipelines stop at object detection or occupancy. The thesis explored whether the same bird’s-eye-view stack could also predict a dense radial-velocity field.

03 / Architecture

Range–Doppler input to joint perception heads.

The pipeline processes range–Doppler input with a hierarchical Swin Transformer, decodes features into range–angle space, and predicts detection, occupancy, and radial velocity through joint heads.

Range–Doppler radar heatmap used as model input — 01 / Range–Doppler inputSignal representation

Hierarchical Swin Transformer feature-extraction stage — 02 / Swin TransformerFeature extraction

Range–angle decoder in the radar perception pipeline — 03 / Range–angle decoderSpatial decoding

Joint detection, occupancy, and velocity prediction heads — 04 / Joint headsDetection · occupancy · velocity

Complete T-FFTRadNet model architecture — Reference architectureT-FFTRadNet

04 / Training

Doppler-aware augmentation across ego speeds.

Training used Doppler-axis shifts to simulate changing ego speed while correcting the velocity targets consistently. The simulation covered ±20.48 m/s, with a maximum shift of ±256 bins and target correction Δv = shift × 0.08 × cos(θ).

Plot of Doppler-bin shift used for velocity-aware training augmentation — Doppler-axis augmentationPhysically consistent target correction

05 / Evidence

Controlled accuracy and highway occupancy improvement.

Airport · controlled evaluation

Low velocity error in a clean, structured scene

Across the evaluated airport speeds, the predicted velocity grid reached an average mean absolute error of approximately 0.21 m/s.

Average velocity MAE≈ 0.21 m/s

Sign accuracyup to 99.99%

Overall accuracyup to 95.88%

Ground-truth and predicted airport velocity grids

Highway · dynamic evaluation

A harder scene with ego motion and moving traffic

The highway sequence was selected as a harder evaluation because ego motion, denser traffic, and moving vehicles create stronger relative velocities and more ambiguous scene structure than the controlled airport sequence.

Ground-truth and predicted highway velocity grids

EvaluationHighway occupancy grid

Confidence threshold≥ 75%

Baseline TPR64.53%

Multi-task TPR67.01%

06 / Scope

What the result demonstrates—and where it remains difficult.

The thesis demonstrates that dense radial-velocity prediction can sit inside the same bird’s-eye-view radar stack as detection and occupancy, adding scene-level motion information without a separate object-level pipeline.

View DirectFlowMatch