DoppDrive: Doppler-Driven Temporal Aggregation for Improved Radar Object Detection

Yuval Haitman^*† and Oded Bialer^*

General Motors, Technical Center Israel
ICCV 2025
^*Indicates Equal Contribution
^†Author is also with the School of Electrical and Computer Engineering in Ben Gurion University of the Negev.

Paper Supplementary Code arXiv

TL;DR

DoppDrive is a Doppler-driven temporal aggregation method for radar point clouds that increases point density and reduces scatter from dynamic objects. This leads to significantly improved object detection—especially at long ranges. Compatible with any detector, DoppDrive is validated across multiple datasets and introduces LRR-Sim, a new 300m-range radar dataset for autonomous driving.

Teaser Figure — **Object detection with different input point clouds:** (1) single frame, highly sparse point cloud, (2) temporally aggregated with ego-motion compensation—scattered points from dynamic objects, (3) Doppler-driven aggregation—dense points with minimal scatter, enhancing object detection.

Comparison of radar point cloud aggregation methods

Left: No aggregation. Middle: Our DoppDrive (Doppler-based dynamic aggregation). Right: Naive aggregation (ego-motion compensation only).

● Blue points: Dynamic radar reflections, ● Orange points: Static radar reflections, ● Gray points: LIDAR points, ▭ Magenta boxes: Ground truth object annotations.
DoppDrive achieves denser and more accurate aggregation of dynamic objects compared to both no aggregation and naive aggregation.

Abstract

Radar-based object detection is essential for autonomous driving due to radar's long detection range. However, the sparsity of radar point clouds, especially at long range, poses challenges for accurate detection. Existing methods increase point density through temporal aggregation with ego-motion compensation, but this approach introduces scatter from dynamic objects, degrading detection performance. We propose DoppDrive, a novel Doppler-Driven temporal aggregation method that enhances radar point cloud density while minimizing scatter. Points from previous frames are shifted radially according to their dynamic Doppler component to eliminate radial scatter, with each point assigned a unique aggregation duration based on its Doppler and angle to minimize tangential scatter. DoppDrive is a point cloud density enhancement step applied before detection, compatible with any detector, and we demonstrate that it significantly improves object detection performance across various detectors and datasets.

Method

Overview of the Doppler-driven temporal aggregation method. (a) Top view of the scenario with the radar-equipped host vehicle and a dynamic vehicle moving toward each other. Points from previous frames \( T_{-1} \) and \( T_{-K} \) are aggregated with the current frame \( T_0 \), compensating for the dynamic vehicle's motion by shifting points along the radial direction based on their dynamic Doppler measurement. (b) Zoom-in on (a), showing the accurate shift \( \boldsymbol{q}^i_k \) vs. the approximated shift \( \tilde{\boldsymbol{q}}^i_k \), with resulting offset distance \( \epsilon^i_k \). (c) Result of our aggregation method, showing denser reflection points with minimal disparity where the aggregation duration of \( \tilde{\boldsymbol{q}}^i_{-K} \) yields an expected offset beyond the tolerated limit, so it is discarded. (d) Result of standard aggregation, showing significant scatter in the aggregated points.

Experimental Results

**Qualitative examples demonstrating DoppDrive in two scenarios:** (a) from the aiMotive dataset (b) from LRR-Sim. Columns show: (left) Sparse point cloud without aggregation, leading to missed detections; (center) Aggregation with ego-motion compensation, causing scattered points around dynamic objects and inaccuracies in object size and position; (right) DoppDrive enhances density around dynamic objects (blue points) with minimal scatter, improving detection accuracy.

Object Detection Accuracy (AP) Across Datasets

Method	aiMotive				Radial				LRR-Sim
Method	SMF	KRD	RPNet	NVR	SMF	KRD	RPNet	NVR	SMF	KRD	RPNet	NVR
No Aggregation	74.6	75.0	73.4	70.9	91.3	91.4	91.0	89.3	89.2	88.7	88.7	86.6
Standard Aggregation	81.7	82.3	80.2	75.2	92.5	92.1	92.3	91.2	91.1	89.9	90.8	90.1
DoppDrive (Ours)	89.1	89.4	87.8	81.8	95.8	95.5	95.7	94.1	93.3	93.1	92.9	92.7

Table: Average Precision (AP) of four detectors: SMURF (SMF), K-Radar (KRD), Radar PillarNet (RPNet), and Nvradarnet (NVR). Results are evaluated across three datasets (aiMotive, Radial, LRR-Sim) using no aggregation, ego-motion-based standard aggregation, and the proposed DoppDrive.

References:
[1] Liu et al., "Spatial Multi-scale Feature Aggregation for Radar Object Detection", CVPR 2023. (SMF)
[2] Paek et al., "Enhanced Radar Object Detection Using Multi-Scale Knowledge Distillation", ICRA 2023. (KRD)
[3] Zheng et al., "RCFusion: Radar-Camera Fusion for Robust Object Detection in All Weather Conditions", ECCV 2022. (RPNet)
[4] Popov et al., "Neural Velocity Radar Network for All-Weather Object Detection", NeurIPS 2023. (NVR)

The LRR-Sim Dataset

LRR-Sim is the first publicly available long-range radar dataset with annotated 3D point clouds extending up to 300 meters, filling a critical gap in radar research for highway and high-speed scenarios. It includes ~25K training and ~5K testing frames across 50 simulated highway scenes, with a total of ~137K annotated vehicles — including ~25K vehicles beyond 175m, where conventional sensors struggle. Scenes were created using the CARLA simulator for realistic world modeling and highway maps, with a high-fidelity radar simulation applied on top to generate 77GHz MIMO radar point clouds. All objects within ±55° azimuth and ±20° elevation are annotated with accurate 3D bounding boxes.

Visit LRR-Sim dataset GitHub for dataset download, technical details, and visualizations.