Abstract:

To address the problem that the limited field of view of single camera 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 multi-camera images. The horizon， as a global geometric prior， can unify these perspective differences， but its detection is easily affected by occlusion and environments， 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. Firstly， perspective transformation and bounding box cropping were applied for data augmentation. Secondly， a diamond space representation was introduced to alleviate instability in vanishing-point learning. Thirdly， Receptive-Field Attention Convolution （RFAConv） and upsampling by Dynamic Sampling （DySample） were used to enhance feature representation and reconstruction accuracy. Finally， a geometric consistency loss function was designed to enhance the constraints of the orientation and position of horizon detection. Experimental results demonstrate that on the BrnoCompSpeed dataset， RoadHoriNet achieves a pixel error of 5.166%， an angle error of 0.032 5°， and a detection accuracy of 94.834%， while reducing the pixel error by 4.815 percentage points and the angle error by 0.019 4° compared with the adaptive horizon detection method. In the task of cross-camera BEV road geometry alignment， the relative alignment accuracy of RoadHoriNet reaches at least 99.129% after being corrected by the RoadHoriNet method， demonstrating its practicality and generalization potential in real-world traffic environments. It can be seen that RoadHoriNet provides a stable geometric prior for camera pose normalization and multi-camera coordinate unification， improving the relative alignment accuracy and robustness of cross-camera BEV road geometric alignment significantly.

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

摘要：

针对高速公路广泛部署的监控相机单相机视野有限，难以实现大范围连续感知的问题，提出跨相机鸟瞰视角（BEV）的道路几何对齐任务，以提升场景一致性与完整性。然而，由于多相机图像间存在视角差异与结构错位，该任务面临较大挑战。地平线作为全局几何先验能够统一视角差异。针对它的检测易受遮挡和环境干扰，制约对齐精度的问题，提出面向跨相机BEV道路对齐的路侧地平线检测方法RoadHoriNet （Road Horizon detection Network）。首先，通过透视变换与包围框裁剪进行数据增强；其次，引入钻石空间表示缓解消失点学习不稳定；再次，结合感受野注意力卷积（RFAConv）与动态上采样（DySample）提升特征表达与重建精度；最后，设计几何一致性损失函数，加强地平线检测的方向与位置约束。在公开数据集BrnoCompSpeed上的实验结果表明，RoadHoriNet 在像素误差（5.166%）和角度误差（0.032 5°）方面表现最佳，检测精度达到94.834%。与自适应地平线检测方法相比，像素误差减小了4.815个百分点，角度误差降低了0.019 4°；在跨相机BEV道路几何对齐任务中，经过RoadHoriNet方法修正后，相对对齐精度达99.129%以上，验证了该方法在交通环境中的实用性与推广潜力。可见，所提出的地平线检测方法能够为相机姿态归一化与坐标统一提供稳定的几何先验，有效提升跨相机BEV道路几何对齐的相对对齐精度与鲁棒性。

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, 动态上采样