多向前方交会与单隐层神经网络结合的近景前方交会法

针对三维坐标求解精度受非线性误差影响的问题,提出了一种多像前方交会与单隐层BP神经网络相结合的方法。基本过程为:①在样本点真实世界坐标已知的条件下,构建关于世界坐标的拉格朗日方程,对相机外参进行优化,以得到更高精度的三维坐标初值。②利用计算得到的三维坐标和真实三维坐标分别作为输入和输出参数对单隐层BP神经网络进行训练。③将三维坐标初值带入网络模型对其进行改正。试验结果表明:①相较于前方交会、稀疏光束法平差及典型神经网络方法,在试验装置环境视场内,本文方法解算精度较高,最大偏差为0.4927 mm。②相较于典型神经网络方法,本文网络结构为3-6-3,结构简单且计算效率高。

Method of close-range space intersection combining multi-image forward intersection with single hidden layer neural network

Aiming at the problem that the three-dimensional coordinate solution precision is influenced by the non-linear error, a method of combining multi-image forward intersection with single hidden layer BP neural network is proposed in this paper. The steps are: ①In order to obtain the initial value of the three dimensional coordinates with higher accuracy, the external parameters of the camera are optimized by constructing the Lagrange equation about the world coordinates under the constraint of known real world coordinates of the sample point. ②The single hidden layer BP neural network is trained by using the calculated 3D coordinate and the real 3D coordinate as input and output parameters, respectively. ③The initial three-dimensional coordinate is corrected by brought into the model. Experiments show that: ①In the environmental field of view of the test device, the proposed method outperforms the space intersection, the sparse bundle adjustment and the other classic neural network methods, the maximum deviation is 0.492 7 mm. ②Compared with other classic neural network methods, the network structure of this paper is 3-6-3, the structure is simple and the calculation efficiency is high.