International Conferences
| SRF: Stereo-Radar Fusion for 3D Object Detection in Adverse Weather Conditions |
| Year |
2026 |
| Month |
|
| Journal |
2026 / IEEE Intelligent Vehicles Symposium (IV) |
| Author |
Batyrbek Mukhatbekov, Dong-Hee Paek, Woo-Jin Jung, and Seung-Hyun Kong* |
|
Existing camera-radar fusion methods for 3D object detection predominantly rely on monocular cameras, which suffer from inherent depth ambiguity, thereby limiting the overall performance. Furthermore, many approaches fuse information in the bird’s-eye-view (BEV) space, discarding crucial 3D information along the elevation axis, or struggle to bridge the representational gap between sparse radar and dense image features. To address these limitations, we propose Stereo-Radar Fusion (SRF), the first end-to-end framework to fuse stereo camera images with raw 4D radar tensor data. The core of our method is a novel Multi-Scale Volume Fusion (MSVF) module. This module integrates a dense, geometrically aware 3D volume generated from a stereo backbone with multi-scale sparse 3D volumes from a 4D radar backbone. By performing fusion directly in voxelized 3D space, our method effectively bridges the modality gap, creating a unified representation that is both dense and metrically accurate. We validate our model on the K-Radar dataset, achieving significant performance improvement for camera-radar fusion. Our method achieves 55.6% 3D mAP on the revision v1.0 and 54.3% on the revision v2.0 versions of the dataset, outperforming the state-of-the-art camera-radar fusion method by 3.9% and 2.3%, respectively. The results demonstrate that utilizing stereo matching properties provides a more accurate 3D representation and superior detection performance, especially in adverse weather conditions. |
