Radar-Coordinate-Modulated State-Space Modeling for mmWave Radar Perception

Abstract

Millimeter-wave (mmWave) radar provides range, azimuth, and radial-velocity measurements, but its range-azimuth-Doppler (RAD) tensors are not uniform image grids. Target evidence and clutter vary with the radar measurement coordinates because range affects attenuation, azimuth affects angular separability, and Doppler is closely related to static clutter and motion ambiguity. Most learned radar backbones still apply largely shared feature-propagation rules across RAD bins. PMMamba addresses this problem by using normalized range, azimuth, and Doppler bin coordinates to modulate the state update and readout of a selective state-space module. To preserve axis-specific context, Tri-Cross-Scan serializes RAD tensors along the three physical axes in both forward and backward directions. For CRUW, which provides temporal sequences of complex range-azimuth maps, time replaces Doppler as the third scan and modulation axis. PMMamba is evaluated on three public radar benchmarks: RADDet for 3-D bounding-box detection, CRUW for range-azimuth radar object detection, and CARRADA for semantic segmentation. It achieves 59.2% mean Average Precision (mAP) on RADDet at Intersection-over-Union (IoU) 0.3, 85.9% Average Precision (AP) on CRUW, and 48.0% mean IoU (mIoU) on CARRADA. Ablation studies on RADDet show that coordinate-conditioned dynamics outperform input-level positional encoding, and that physical-axis scanning is more effective than flat raster scanning for RAD tensors. The results show that radar-bin coordinates are more useful when they modulate the state update and readout than when they are only added to the input features. The source code is publicly available at https://github.com/Hl-Chen9/PMMamba