关键词: 地平线检测, 钻石空间, 几何一致性损失, 多相机对齐, 感受野注意力卷积（RFAConv）, 动态上采样（DySample）

Abstract: Abstract: To address the problem that the limited field of view of widely deployed highway surveillance cameras makes it difficult to achieve large-scale continuous perception, a cross-camera Bird’s-Eye-View (BEV) road geometric alignment task was proposed to improve scene consistency and completeness. However, this task faces challenges due to the perspective differences and structural misalignments among multiple cameras. The horizon, as a global geometric prior, can unify these perspective differences, but its detection is easily affected by occlusion and environmental interference, which becomes a key factor limiting alignment accuracy. To solve this problem, a horizon detection method for cross-camera BEV road alignment, named RoadHoriNet (Road Horizon detection Network), was proposed as a key technology. First, perspective transformation and bounding box cropping were applied for data augmentation. Then, a diamond space representation was introduced to alleviate instability in vanishing-point learning. Next, Receptive Field Attention Convolution (RFAConv) and Dynamic Upsampling (DySample) were used to enhance feature representation and reconstruction accuracy. Finally, a geometric consistency loss function was designed to constrain the orientation and position of the detected horizon.Experimental results demonstrated that RoadHoriNet achieved a pixel error of 5.166%, an angle error of 0.0325°, and a detection accuracy of 94.834% on the BrnoCompSpeed dataset, reducing the pixel error by 4.815 percentage points and the angle error by 0.0194° compared with the adaptive horizon detection method. In the cross-camera BEV road alignment task, the overall alignment accuracy reached 99.129%.It is concluded that RoadHoriNet provides a stable geometric prior for camera pose normalization and multi-camera coordinate unification, significantly improving the accuracy and robustness of cross-camera BEV road geometric alignment.

Key words: horizon detection, diamond space, geometric consistency loss, multi-camera alignment, Receptive Field Attention Convolution (RFAConv), Dynamic Upsampling(DySample)