.png){kind=icon}
Return
Weakly-supervised multi-view BEV mapping with point pair optimization
.png){kind=icon}
PRE
AI
DOI：10.1080/13682199.2026.2641967.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
To address the challenges of low mapping accuracy and target occlusion in multi-view roadside BEV perception, this paper proposes a weakly supervised bird's-eye view mapping framework. The method employs independent Transformer encoders to process images from each viewpoint and achieves accurate BEV alignment without per-view correspondence labels through cross-view consistency constraints, eliminating the need for dense BEV annotations or precise camera calibration. Experiments on the DAIR-V2X and a proprietary multi-view intersection dataset demonstrate that the framework outperforms homography-based mapping, BEVDet, and BEVFormer in terms of Point Mapping Squared Error (PMSE) and Overlap Rate (OR). Ablation studies further validate the effectiveness of the pairing loss and the independent mapping strategy. Although inference speed is slightly lower than homography-based methods, the proposed approach offers a practical balance between accuracy and efficiency for real-time roadside perception tasks.
Keywords:
Bird's-eye view (BEV) mapping
multi-view detection
transformer encoding
point matching optimization
roadside perception
Journal
I
IF:
1.1
Papers:
61
Citations:
0
Organization
No organization information available
Cited Papers
No cited papers available