4D Radar Perception Algorithms for Autonomous Driving: A Review

Summary

Abstract Summary

Research on 4D millimeter-wave radar perception algorithms has flourished in recent years, extending from signal processing and object detection to semantic segmentation, motion estimation, occupancy prediction, and dynamic scene reconstruction. This review organizes the field according to the evolution of perception tasks and algorithms. It first introduces radar fundamentals, data representations, and quality-enhancement methods, and then reviews object-level perception, motion and localization, local and dense spatial perception, and dynamic scene understanding. Across these directions, we compare radar-only learning, multimodal fusion, and cross-modal supervision and knowledge distillation. Particular attention is paid to how elevation, Doppler measurements, and radar physical priors are exploited across tasks. We further summarize the task coverage, input data, annotations, and evaluation protocols of existing datasets, clarifying the empirical support for different research directions. Finally, we discuss the common challenges and future directions of 4D radar perception for autonomous driving. This review provides a task-oriented perspective on the transition from sparse object perception to dynamic spatial understanding.

Key Points

Research on 4D millimeter-wave radar perception algorithms has flourished in recent years, extending from signal processing and object detection to semantic segmentation, motion estimation, occupancy prediction, and dynamic scene reconstruction.
This review organizes the field according to the evolution of perception tasks and algorithms
It first introduces radar fundamentals, data representations, and quality-enhancement methods, and then reviews object-level perception, motion and localization, local and dense spatial perception, and dynamic scene understanding.
Across these directions, we compare radar-only learning, multimodal fusion, and cross-modal supervision and knowledge distillation.
Particular attention is paid to how elevation, Doppler measurements, and radar physical priors are exploited across tasks.

Radar Point-Cloud and Tensor Enhancement

Radar point-cloud and tensor enhancement is a moderately mature direction. As summarized in Fig. 3 , 4D radar point clouds are typically sparser and less geometrically complete than LiDAR point clouds, but they offer longer detection range, direct radial-velocity and RCS measurements, and stronger adverse-weather robustness [ 1 , 51 , 55 ] .

This complementarity motivates enhancement that improves geometric usability without discarding radar-specific motion and scattering evidence.

Radar Point-Cloud and Tensor Enhancement — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 4: Overall architecture of CORENet for LiDAR-supervised 4D radar denoising and object detection [ 5 ] .

Object-Level Perception

Object-level perception is one of the most mature and extensively studied 4D radar directions. Research has progressed from early radar object classification [ 83 ] to extensive 3D detection, with practical baselines established for both radar-only and fusion-based perception. Current work continues to improve recognition of vulnerable road users under sparse returns [ 86 ] .

Current object-level studies follow three technical routes: radar-only learning, multimodal fusion, and cross-modal supervision.

Object-Level Perception — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 5: Overall two-stage architecture of CVFusion for camera–4D-radar 3D object detection [ 16 ] .

Object Motion Perception

Motion perception comprises subdirections with unequal maturity. Multi-object and extended-target tracking form the most established branch. Moving-instance segmentation, ego-motion estimation, odometry, localization, and SLAM have reached an intermediate stage, whereas scene-flow estimation remains an early but active direction. Object tracking is studied mainly through radar-only tracking and camera–radar fusion.

Radar-only methods follow tracking-by-detection, extended-object filtering, or box-free moving-instance tracking.

Object Motion Perception — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 6: Self-supervised scene-flow and motion-segmentation framework of RadarSFEMOS. Correlation features are constructed from adjacent 4D radar point clouds; coarse scene flow is first estimated and then refined by a diffusion model, while rigid-motion residuals are used to segment moving objects [ 25 ] .

Local Spatial Geometry and Depth Completion

Radar-guided depth completion and radar-based road-boundary perception remain early-stage 4D radar directions. Current work follows two main routes: camera–radar methods infer dense image-aligned metric depth from sparse radar ranges, whereas radar-only methods estimate local road-boundary geometry from temporal point clouds.

Radar already measures range at observed returns; depth completion therefore fills unobserved image regions rather than re-estimating the measured radar range.

Local Spatial Geometry and Depth Completion — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 7: Radar occupancy learning framework of 4D-ROLLS. LiDAR rays and height maps generate occupancy pseudo-labels; a sparse-to-dense backbone trains a radar-only occupancy network that jointly outputs occupancy prediction and BEV segmentation [ 10 ] .

Dense Spatial Perception

Within dense spatial perception, point-wise semantic segmentation remains an early-stage direction, whereas grid- and voxel-based occupancy prediction has developed rapidly into a major 4D radar task. Point-wise segmentation assigns semantic labels only to measured radar returns, while occupancy prediction estimates free, occupied, and semantic states on a regular 3D grid, including locations without a return.

Existing point-wise segmentation studies are dominated by cross-modal knowledge distillation.

Dense Spatial Perception — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 8: Conceptual levels of dynamic dense scene understanding. Scene flow estimates sparse point-wise 3D motion, dynamic occupancy would associate voxel states with motion, and dynamic reconstruction maintains cross-frame trajectories and scene geometry.

Dynamic Dense Scene Understanding

Dynamic dense scene understanding is an exploratory direction that aims to represent both spatial structure and temporal evolution. Existing studies provide partial solutions at three levels—point-wise motion, temporally updated voxel states, and cross-frame scene reconstruction. However, a unified representation connecting these outputs has not yet emerged. Figure 8 summarizes these three conceptual levels.

At the point level, scene-flow methods recover the 3D motion of sparse radar returns between adjacent frames.

Dynamic Dense Scene Understanding — 4D Radar Perception Algorithms for Autonomous Driving: A Review — Fig. 9: Representative radar data forms and annotation strategies. Panels (a) and (b) show camera and range-Doppler frames from I/Q-1M and point-wise semantic labels from ZJUSSet, respectively [ 3 , 31 ] . Panels (c) and (d) show VoD-derived radar-LiDAR scene-flow labels and synchronized adverse-condition modalities from DSERT-RoLL, respectively [ 26 , 40 ] . Together, the panels contrast direct measurements with point-wise, generated-motion, and multimodal annotation strategies.

Datasets and Benchmark Protocols

Early 4D radar datasets concentrated on object-level perception. VoD and TJ4DRadSet established point-cloud benchmarks for 3D detection, while K-Radar added radar tensors and adverse-weather sequences. RaDelft subsequently provided lower-level radar data for automotive radar detection [ 80 ] . Task coverage later expanded beyond detection.

ZJUSSet provides point-wise semantic labels, and the Radar-LiDAR Scene Flow Dataset repurposes synchronized VoD sequences with generated motion labels.

Challenges and Future Directions

First, measurement-level automotive radar data remain scarce because many road-driving datasets release processed point clouds rather than measurement-level I/Q data.

Future datasets should jointly release raw I/Q, radar tensors, point clouds, BEV products, and synchronized auxiliary sensors together with the calibration, synchronization, waveform, array, CFAR, and point-cloud-generation parameters needed to reproduce front-end processing. Second, cross-modal supervision should not be limited to LiDAR-like geometric completion.

Sources and Demonstrations

Figures are reproduced from Wu et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.