-
实验中相机选用德国Balser ace acA2500-20 gm,分辨率为2590 pixel×2 048 pixel,水平像元尺寸与垂直像元尺寸均为4.8 μm,镜头焦距为12 mm,相机的内参以及畸变参数由相机标定得到[21]。使用12个固定参考点,通过相机拍摄参考点图像,实验装置如图5所示,计算机配置与仿真实验相同。
随机选取两个参考点添加噪声模拟粗差。通过4种算法解算得到位姿参数,计算参考点的像平面重投影结果,通过比较重投影结果与实测值来反映位姿解算精度[22]。由于WOI与WAOI精度相近,重投影图中不易分辨,因此不包含WOI。参考点在像平面上的重投影结果如图6所示。
从图6可以看出,WAOI对应的解算结果与测量结果非常接近,而SRPnP和OI的解算结果与测量结果有较大的偏差。进一步分析各方法测量精度差异,计算12个参考点重投影位置如表1所示,其中x、y分别为参考点在像平面像素坐标,r为参考点重投影结果与测量结果之间的距离。
Reference
pointMeasured SRPnP calculated OI calculated WAOI calculated x y xs ys rs xo yo ro xw yw rw 1 1112.04 1075.67 1108.25 1095.76 20.44 1107.75 1097.65 22.39 1111.39 1076.11 0.78 2 962.89 1213.99 956.77 1222.14 10.19 957.69 1223.6 10.93 962.52 1214.76 0.86 3 1104.29 1144.72 1104.88 1144.68 0.59 1105.66 1143.59 1.78 1104.67 1144.76 0.39 4 1183.2 937.3 1188.71 945.64 9.99 1187.18 945.64 9.24 1182.99 937.53 0.31 5 1036.74 1144.08 1032.25 1155.25 12.04 1032.4 1156.62 13.27 1036.71 1144.02 0.07 6 1029.16 1208.41 1029.15 1199.62 8.79 1030.57 1198.03 10.48 1029.72 1208.45 0.56 7 1242.78 1218.83 1243.26 1220.98 2.2 1245.74 1217.98 3.07 1241.91 1219.16 0.93 8 1174.45 1077.53 1179.45 1074.51 5.84 1179.88 1072.5 7.4 1174.67 1076.73 0.83 9 1181.95 1007.05 1185.44 1019.91 13.33 1184.75 1020.2 13.45 1182.32 1007.41 0.52 10 966.16 1069.77 963.33 1070.24 2.87 962.12 1071.32 4.33 966.38 1070.43 0.7 11 971.22 929.21 967.57 929.76 3.69 963.96 931.69 7.67 971.2 928.51 0.7 12 1179.21 1148.97 1177.77 1162.65 13.75 1178.8 1162.49 13.53 1178.89 1148.63 0.47 Table 1. Image plane projection results comparison (Unit:pixel)
从表1可以看出,WAOI解算得到的参考点重投影均方根误差为0.64 pixel,SRPnP和OI分别为10.29 pixel和11.18 pixel。WAOI与SRPnP和OI相比,参考点重投影结果与测量结果之间的距离要低一个数量级,表明WAOI的位姿解算精度更高。此外解算时间方面,OI平均耗时为23.64 ms,WOI平均耗时为31.89 ms,WAOI平均耗时为8.02 ms,实验表明,WAOI与OI和WOI相比,运算效率分别提升了66.07%和74.85%,具有明显优势。
Pose estimation of camera based on weighted accelerated orthogonal iterative algorithm
doi: 10.3788/IRLA20220030
- Received Date: 2022-03-01
- Rev Recd Date: 2022-04-06
- Available Online: 2022-11-02
- Publish Date: 2022-10-28
-
Key words:
- machine vision /
- pose estimation /
- weighted orthogonal iterative /
- adaptive weights
Abstract: Pose estimation in monocular vision is a key problem in three-dimensional measurement, which is widely used in machine vision, precision measurement and so on. This problem can be solved by n-point perspective (PnP) algorithm. Orthogonal iterative algorithm (OI), as the representative of PnP algorithm, has been widely used in practice because of its high precision. In order to further improve the robustness and computational efficiency of OI algorithm, a weighted accelerated orthogonal iterative algorithm (WAOI) is proposed in this paper. Firstly, the weighted orthogonal iterative algorithm is deduced according to the classical orthogonal iterative algorithm. The weighted collinearity error function is constructed and the weight is updated by using the object point reprojection error to achieve the purpose of iteratively optimizing the pose estimation results. Secondly on this basis, through adaptive weights, the calculation of translation vector and objective function in each iteration is integrated to reduce the amount of calculation in the iterative process, so as to accelerate the algorithm. The experimental results show that when there are two rough points in the 12 reference points, the reprojection accuracy of the reference point of WAOI is 0.64 pixel, the operation time is 8.02 ms, the accuracy is high and the running speed is fast, so it has strong engineering practical value.