单目视觉下道路场景车辆空间形态估算

Vehicle spatial morphology estimation in road scene under monocular camera

针对受射影几何学的限制,单目相机无法直接获得准确的三维点云数据及目标尺度信息,难以获取目标的三维结构问题,提出了一种基于单目交通相机的车辆空间形态估算方法。首先建立道路场景的自动标定模型以获取3D-2D的投影映射及尺度信息,并基于“钻石空间”方法,利用统计轨迹直线及车辆边缘精确求取场景中的地平线,根据标定信息及灭点约束共同构建车辆空间形态的几何约束模型,然后在图像中提取车辆的实际投影约束,包括基于获得的车辆序列轮廓约束,及车辆自身边缘约束,并据此构建误差约束函数,估计车辆空间形态的投影误差,最后根据车辆的初始识别结果及先验信息,优化参数约束空间,并利用误差约束函数在约束空间中迭代求最优,得到精确的车辆空间形态信息。利用公开数据集BrnoCompSpeed及实际道路采集的视频数据共同验证该算法,并与现有类似算法进行比较。结果表明:该算法对于道路场景的适应性强,所需先验条件少,对于多种类型车辆在三维尺寸的估计精度高达94%以上。同时,该算法还可估算车辆实时的空间位置及相对于路面的偏转角度,综合空间形态估算的精度达到92%以上,且实时性较好,单帧多车的估算时间小于0.5 s。与现有算法相比,该方法更适合在道路场景中利用固有的监控相机识别车辆空间形态。

Due to the limitation of projective geometry,it was difficult for monocular camera to obtain accurate 3D point cloud and scale for the three-dimensional structure of the object.To solve this problem,a method of vehicle spatial morphology estimation based on monocular traffic camera was proposed.Firstly,an automatic calibration model of the road scene was established to obtain the 3D-2D projection mapping and scale information.Based on the“diamond space”method,the horizon line of the scene could be accurately obtained by vehicle trajectories and edges.Then,the geometric model of the vehicle spatial morphology could be jointly constructed with calibration information and vanishing point constraints.Secondly,the projection constraints of the vehicle were extracted from the image,including sequences of vehicle contour constraints and vehicle edge constraints.Based on these constraints,the error constraint function could be derived to estimate the projection errors of the vehicle spatial morphology.Finally,according to the initial vehicle recognition results and prior information,the parameter constraint space could be iteratively optimized according to the error constraint function,and the accurate vehicle spatial morphology information could be obtained.The experiments were validated on the public dataset BrnoCompSpeed and videos were collected from actual roads.The proposed method was also compared with similar methods.The results show that the proposed method is strongly adaptive to various road scenes with an accuracy of more than 94%for 3D vehicle size estimation,which requires few prior conditions.In the meanwhile,real-time vehicle spatial position and deflection angle relative to the road plane can be estimated with a comprehensive accuracy of more than 92%for vehicle spatial morphology estimation and a process speed of less than 0.5 seconds for a single frame with several vehicles.Moreover,compared with existing methods,the proposed method is more suitable for vehicle spatial morphology estimation by using surveillance cameras in road scenes.8 tabs,10 figs,35 refs.