ECPFusion: BEV-Based Collaborative Enhancement Fusion for Multi-Modal 3D Object Detection
2026-99-2422
11/20/2026
- Content
- Advancements in autonomous driving technology have increased the requirements for the detection capabilities of environmental perception systems. Multi-sensor fusion has gradually become an important research direction for 3D object detection. Compared with the fusion method using LiDAR, the fusion of cameras and millimeter-wave radars has certain advantages in terms of cost and environmental adaptability. However, in the actual multimodal fusion process, the existing 3D object detection methods for radar-camera mainly focus on the alignment mode of the modalities, ignoring the local response offset generated during the pre-view transformation of BEV fusion, which causes instability in spatial representation and the limitation on further improvement of BEV representation quality and 3D detection performance due to the accumulation of redundant features after fusion. To address these obstacles, Therefore, This article suggests a refined fusion architecture, ECPFusion. This method introduces a multi-scale feature enhancement module (FCBM) before fusion to alleviate the view transformation error, and designs a fusion feature refinement module (REFFM) after fusion to suppress redundant features and strengthen key semantic expression. Experimental results show that on the VoD and TJ4DRadSet datasets, this method achieves 3D mAP of 57.10% and 38.85% respectively, verifying its effectiveness in cases of driverless car operations.
- Citation
- Li, X. and Huang, H., "ECPFusion: BEV-Based Collaborative Enhancement Fusion for Multi-Modal 3D Object Detection," 2026 2nd International Conference on Intelligent Transportation and Future Mobility (ITFM2026), Xi'An, China, March 27, 2026, https://doi.org/10.4271/2026-99-2422.