Enhancing Camera-Radar Perception Via Explicit BEV Alignment and Fusion
Year: 2026
Type: conference-paper
Abstract: 3D object detection is a fundamental task in autonomous driving perception where cameras and millimeter-wave radars are commonly combined to provide complementary semantic and geometric information. However, depth prediction errors in camera-based Bird's-Eye-View generation often introduce spatial misalignment. This issue increases the burden of implicit geometric correction in query-based detection decoders and limits the effectiveness of multi-modal integration. To alleviate this issue, we propose a Hybrid BEV Alignment and Fusion (H-BAF) module placed before the Transformer decoder. This module performs cross-modal spatial alignment and image feature enhancement to produce BEV representations with improved spatial consistency. The proposed method explicitly calibrates spatial correspondences and utilizes radar geometric cues to strengthen image features. Experiments on the nuScenes validation set demonstrate competitive performance in terms of NDS and mAP. These results validate the effectiveness of the proposed H-BAF module for robust multi-modal 3D object detection.
Authors: Sin-Ye Jhong, Yen-Yu Hsieh, Ping-Han Chen
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204148205