UGBCF-Net: Uncertainty-Guided BEV Cross-Attention Fusion Network for Robust Radar-Camera Object Detection

Engineering notes

Key topics: autonomous driving system, autonomous driving, bev, object detection, nuscenes, radar. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

Object detection under adverse weather remains a major challenge for autonomous driving systems. Cameras do not work well during rain, fog, and darkness, while mmWave radar works accurately in all of these conditions, regardless of illumination and precipitation. In this paper, we propose UGBCF-Net – the Uncertainty-Guided Bird's-Eye-View Cross-Attention Fusion Network, which incorporates radar confidence estimation based on physical laws and adaptive multimodal fusion. Our proposed architecture estimates radar confidence from radar cross-section measurements, maintains the Doppler channel by means of four-channel BEV representation, and uses uncertainty-guided weighting with a lightweight pooled cross-attention network for efficient feature fusion. Experimental results on the nuScenes v1.0-trainval dataset have shown that our UGBCF-Net reaches 0.683 mAP with 8.9 M parameters, operating at 56 FPS on a single NVIDIA RTX 3050 GPU. UGBCF-Net has achieved an accuracy similar to large-fusion networks and showed great robustness, reaching night recall equal to 0.884 with recall rates of 93.1%, 95.9% and 97.4% for vehicles, pedestrians, and cyclists, respectively.

Links and sources

• PDF from original source

Comments