DRIFT: Dual-Representation Inter-Fusion Transformer for Automated Driving Perception with 4D Radar Point Clouds

Siqi Pei; Andras Palffy; Dariu M. Gavrila

cs.ROMar 10, 2026

DRIFT: Dual-Representation Inter-Fusion Transformer for Automated Driving Perception with 4D Radar Point Clouds

Authors: Siqi Pei, Andras Palffy, Dariu M. Gavrila

Organizations: Delft University of Technology, Delft, The Netherlands · Perciv AI, Delft, The Netherlands

Abstract

4D radars, which provide 3D point cloud data along with Doppler velocity, are attractive components of modern automated driving systems due to their low cost and robustness under adverse weather conditions. However, they provide a significantly lower point cloud density than LiDAR sensors. This makes it important to exploit not only local but also global contextual scene information. This paper proposes DRIFT, a model that effectively captures and fuses both local and global contexts through a dual-path architecture. The model incorporates a point path to aggregate fine-grained local features and a pillar path to encode coarse-grained global features. These two parallel paths are intertwined via novel feature-sharing layers at multiple stages, enabling full utilization of both representations. DRIFT is evaluated on the widely used View-of-Delft (VoD) dataset and a proprietary internal dataset. It outperforms the baselines on the tasks of object detection and/or free road estimation. For example, DRIFT achieves a mean average precision (mAP) of 52.6% (compared to, say, 45.4% of CenterPoint) on the VoD dataset.

Figures & tables

Fig. 1 : Example scene from the View-of-Delft [ 1 ] dataset with radar and annotation data. The two inlets on the left show corresponding LiDAR and radar point clouds around the same pedestrian. Unlike the LiDAR point cloud, it is difficult to detect the pedestrian using only the local region of the radar point cloud due to its sparsity. However, incorporating global information (i.e. the pedestrian’s relative position to the ego-vehicle and the scene, e.g. drivable area marked with blue) with local features such as shape and velocity, the pedestrian’s presence becomes more apparent.

Panel of Figure 2 — Fig. 2 : Comparison of dual-representation model structures. FS denotes feature sharing block. (a) The sequential path processes the point cloud through a point-based path first, then through a voxel-based path, e.g. [ 4 ] . (b) The parallel path with fusion at the end processes both paths independently and merges their outputs before proceeding with a single-representation path, e.g. [ 5 , 6 ] . (c) Proposed structure processes both paths in parallel and introduces feature sharing at each intermediate stage between them.

Fig. 3 : Model architecture. The top figure shows the overall architecture of DRIFT, which consists of a point path, a pillar path, and feature sharing blocks. The feature after block

i

M_{i}\times C_{i}\ \text{in}\ H_{i}\times W_{i}

refers to a sparse pillar representation tensor with shape

M_{i}\times C_{i}

, from BEV grid size

H_{i}\times W_{i}

, where

M_{i}

is the number of non-empty pillars and

C_{i}

is the number of channels. (a) Point transformer block. (b) Pillar transformer block. (c1) Feature sharing block with add or concatenation fusion. (c2) Feature sharing block with cross-attention fusion.

Method	Entire Area				Driving Corridor
Method	mAP $\uparrow$	Car $\uparrow$	Ped. $\uparrow$	Cyc. $\uparrow$	mAP $\uparrow$	Car $\uparrow$	Ped. $\uparrow$	Cyc. $\uparrow$
PointPillars † [ 9 ]	45.2	37.1	35.0	63.4	67.5	70.2	47.2	85.1
CenterPoint † [ 2 ]	45.4	32.7	38.0	65.5	65.1	62.0	48.2	85.5
CenterPoint * [ 2 ]	47.4	39.9	37.8	64.7	67.4	71.3	47.3	83.7
PillarNeXt [ 29 ]	42.2	30.8	33.1	62.8	63.6	66.7	39.0	85.1
RadarPillars [ 12 ]	50.7	41.1	38.6	72.6	70.5	71.1	52.3	87.9

TABLE I : Comparison of VoD validation set results. DRIFT uses cross-attention feature sharing. Pre-training is performed by perciv-scenes-2 dataset. The best results of each configuration (with and without pre-training) are bold. Results marked with † are inherited from [ 6 ] , results marked with * comes from re-implemented methods.

Method	Object Detection					Free-road
Method	mAP $\uparrow$	NDS $\uparrow$	$\text{AP}_{\text{car}}$ $\uparrow$	$\text{AP}_{\text{ped.}}$ $\uparrow$	$\text{AP}_{\text{cyc.}}$ $\uparrow$	$\text{IoU}_{\text{free}}$ $\uparrow$	$\text{IoU}_{\text{occ}}$ $\uparrow$
CenterPoint [ 2 ]	51.8	50.9	55.5	45.2	54.6	71.5	28.6
DRIFT (attention)	55.2	52.7	56.0	51.8	57.8	73.3	29.7

TABLE II : Comparison of object detection and free-road segmentation results on perciv-scenes-2 validation set.

Panel of Figure 6 — Fig. 4 : Visualization results of ground truth, CenterPoint, and DRIFT (proposed) on the VoD validation set. Blue/green/red boxes indicate car, pedestrian, and cyclist detections, respectively. All images are trimmed.

Panel of Figure 7 — Fig. 5 : Visualization of free-road ground truth and prediction on perciv-scenes-2 validation set on BEV. Yellow segment indicate free-road region.

Configuration	$\text{mAP}_{\text{all}}$ $\uparrow$	$\text{mAP}_{\text{roi}}$ $\uparrow$
Pillar Path (no trans.)	46.8 $\pm$ 0.6	67.5 $\pm$ 0.1
Pillar Path (no trans., double channel)	47.5 $\pm$ 0.6	68.1 $\pm$ 2.2
Pillar Path (trans. in 1st block)	48.6 $\pm$ 1.3	69.3 $\pm$ 1.0
Pillar Path (trans. in all blocks)	49.6 $\pm$ 0.7	69.3 $\pm$ 0.8
Pillar Path (trans. in all blocks, double channel)	50.0 $\pm$ 0.5	69.1 $\pm$ 0.8
Dual Path (trans. in 1st blocks)	49.2 $\pm$ 0.8	68.9 $\pm$ 0.5

TABLE III : Ablation study on the two data representation paths and transformer blocks.

Fusion Configuration		$\text{mAP}_{\text{all}}$ $\uparrow$	$\text{mAP}_{\text{roi}}$ $\uparrow$
Point to Pillar	Pillar to Point	$\text{mAP}_{\text{all}}$ $\uparrow$	$\text{mAP}_{\text{roi}}$ $\uparrow$
Add	Add	49.9 $\pm$ 0.7	69.7 $\pm$ 0.5
Concat	Concat	49.4 $\pm$ 0.9	69.8 $\pm$ 0.8
Attention	Attention	51.5 $\pm$ 1.1	70.6 $\pm$ 0.8
Add	Attention	50.7 $\pm$ 0.2	69.6 $\pm$ 1.1
Attention	Add	50.2 $\pm$ 1.2	69.6 $\pm$ 0.6

TABLE IV : Comparison of different types of feature sharing methods.

Fusion Configuration	$\text{mAP}_{\text{all}}$ $\uparrow$	$\text{mAP}_{\text{roi}}$ $\uparrow$
None (pillar path only)	49.0 $\pm$ 0.4	69.4 $\pm$ 0.6
Point to Pillar Only	49.4 $\pm$ 0.2	68.8 $\pm$ 0.7
First Block Only	49.4 $\pm$ 0.4	69.9 $\pm$ 0.4
Last Block Only	49.9 $\pm$ 0.2	69.1 $\pm$ 0.2
All Blocks	51.5 $\pm$ 1.1	70.6 $\pm$ 0.8

TABLE V : Ablation study on feature sharing blocks.

Autonomous Driving Perception

Autonomous driving perception enables vehicles to detect and interpret road users, drivable space, and hazards using onboard and shared sensors.

See all 379 papers

Accuracy- and Real-Time-Aware 4D Radar Preprocessing for Autonomous Driving Perception Systems
Sep 16, 2026Woo-Jin Jung, Dong-Hee Paek, Jeong-Su Park +1Autonomous Driving Perception 3D Object Detection
Cho Chun Shik Graduate School of Mobility Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34051, Republic of Korea · Mechanical Engineering Research Institute Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34141, Republic of Korea · Autonomous Driving Perception Technology Vanguard Team Hyundai Motor Company Seongnam-si 13529, Republic of Korea
RCGDet3D: Rethinking 4D Radar-Camera Fusion-based 3D Object Detection with Enhanced Radar Feature Encoding
May 20, 2026Weiyi Xiong, Bing ZhuCamera-Radar Fusion 3D Object Detection
School of Automation Science and Electrical Engineering, Beihang University, Beijing, P.R. China
R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection
Mar 12, 2026Zhongyu Xia, Yousen Tang, Yongtao Wang +23D Object Detection Camera-Radar Fusion
Wangxuan Institute of Computer Technology, Peking University · EBTech Co. Ltd

Sep 16, 2026J/K move · Enter open · S save

Accuracy- and Real-Time-Aware 4D Radar Preprocessing for Autonomous Driving Perception Systems

Woo-Jin Jung, Dong-Hee Paek, Jeong-Su Park, Seung-Hyun Kong

Cho Chun Shik Graduate School of Mobility Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34051, Republic of Korea · Mechanical Engineering Research Institute Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34141, Republic of Korea · Autonomous Driving Perception Technology Vanguard Team Hyundai Motor Company Seongnam-si 13529, Republic of Korea

Autonomous Driving Perception 3D Object Detection Real-Time Object Detection Millimeter-Wave Radar

Open paper AskPDF arXiv

4D radar has emerged as a promising next-generation sensor for improving the robustness of autonomous driving perception systems because of its stable sensing capability under adverse weather conditions. However, deploying 4D radar in embedded environments with limited hardware resources requires radar-representation preprocessing that jointly considers perception accuracy, real-time performance, and computational complexity. This paper proposes a preprocessing framework for 4D-radar-based 3D object detection. First, Percentile-based 3D Shape Preservation (P3DP) extracts point clouds from radar tensors while preserving object-shape information and suppressing noise and false alarms. Second, Multi-frame-based Noise Point Discrimination using Kernel Density Estimation (MF-KDE) improves the density and reliability of sparse radar point clouds. Finally, Embedded & NetScore (ENS) evaluates suitability for embedded deployment by jointly considering accuracy, real-time performance, adverse-weather robustness, and model complexity.

Abstract

Autonomous Driving Perception

Explore similar work

Accuracy- and Real-Time-Aware 4D Radar Preprocessing for Autonomous Driving Perception Systems

RCGDet3D: Rethinking 4D Radar-Camera Fusion-based 3D Object Detection with Enhanced Radar Feature Encoding

R4Det: 4D Radar-Camera Fusion for High-Performance 3D Object Detection