高速公路相机自动标定及道路坐标系构建

为解决高速公路场景下利用视频监控系统正确描述车辆相对于道路的空间位置问题，通过引入Frenet坐标系概念，提出一种基于相机自动标定的道路坐标系模型。在相机自动标定阶段，利用线分段拟合方法从曲线车辆轨迹中提取平行于直线路段的轨迹点，并通过级联霍夫变换精确估计道路方向的消失点。然后,根据多车辆三维模型约束，对相机参数进行迭代优化。基于标定结果，将车辆轨迹映射到世界坐标系平面上，并用3次样条插值进行拟合。根据大量运动车辆在道路平面内形成的轨迹域分布特征，综合边界约束估计道路中心点。最后，结合道路中心线在各点处的法线向量与车道宽度信息确定平移量，并利用点平移运动拟合车道线，实现道路坐标系的自动建立。使用真实高速公路视频数据，在多种道路条件下进行试验。研究结果表明:在标定阶段，构建方法对不同高速公路场景的最大标定误差不超过11.55%;与最新的方法相比，直线道路平均标定误差分别降低6.68%和3.58%，弯曲道路平均标定误差分别降低7.43%和2.61%;在道路坐标系构建阶段，构建方法的平均投影距离为0.077 m，接近最新方法的0.069 5 m；而其平均精度为0.916，显著优于最新方法的0.663；所提道路坐标系能够自适应道路形态的变化，有效解决了从监控视频中描述车辆与道路之间相对位置关系的问题。

Automatic Camera Calibration and Road Coordinate System Construction in Highways

To accurately describe the spatial position of vehicles relative to the road using video surveillance systems on highways, a road coordinate system construction model based on automatic camera calibration is proposed by introducing the concept of the frenet coordinate system. In the automatic camera calibration stage, the line-segment fitting algorithm was leveraged to extract the trajectory points parallel to the straight road section from the curved vehicle trajectories, and the vanishing point in the road direction was accurately estimated using the cascaded Hough transform. Thereafter, the camera parameters were optimized iteratively according to the multivehicle three-dimensional model constraints. Based on the calibration results, the vehicle trajectories were mapped on the world coordinate system plane and fitted using cubic spline interpolation. According to the distribution characteristics of the trajectory domain formed by a large number of moving vehicles on the road plane, the road center points were estimated using the boundary constraints. Finally, the translation vector was determined by combining the normal vector of the road central line at each point with the lane width information, and the lane line was fitted via the point translation movement to automatically establish the road coordinate system. The experiments were conducted under various road conditions using real highway video data. The experimental results demonstrate that in the calibration stage, the maximum calibration errors of this study for different highway scenes do not exceed 11.55%. In comparison with the latest methods, the average calibration errors of straight roads are reduced by 6.68% and 3.58%, whereas the average calibration errors of curved roads are reduced by 7.43% and 2.61%, respectively. In the road coordinate system construction stage, the average projection distance of the proposed lane line-fitting algorithm is 0.077 m, which is close to that of the latest method (0.069 5 m). The average accuracy of the proposed lane line-fitting algorithm is 0.916, which is significantly better than that of the latest method (0.663). In addition, the proposed road coordinate system can self-adapt to changes in road morphology, effectively solving the problem of describing the relative position relationship between vehicles and roads in surveillance videos.