Study Of Non-metric Digital Camera Calibration And Multi-baseline Bundle Adjustment

With the development of non-metric camera manufacturing technology,the digital camera has been widely used in close-range photogrammetry.Due to the non-metric camera has the advantage of convenient carrying,high degree of popularity and getting abundant information of the target without touching the object,it has been playing an important role in the deformation monitoring of slope filed,the reconstruction of ancient buildings,and archaeological field.This paper discussed the imaging principle of the non-metric camera,and developed an algorithm applicable to the non-metric camera after advanced studying the analytic method in close-range photogrammetry.However,there are large distortion of the objective lens and the distortion of the principal point coordinates of the non-metric camera,it is necessary to establish the corresponding model to eliminate or weaken errors.This paper used the BP neural network model and the improved ten parameter model to calibrate the internal parameters of non-metric digital camera,discussed and analysed the calibration accuracy of the two models through the experiment;Considering the characteristics of large system error in non-metric camera,this paper regards the distortion parameters of the lens as additional parameters,and regards coordinates of the image points as the observation values,integrally calculating the corrected value that coordinates of the corresponding object and the parameters of the camera by the strict bundle adjustment;In addition,a self-adaptive iteration algorithm is proposed in this paper to solve the ill-conditioned normal equation.During the experiment,two-dimensional testing ground is established firstly,then calculating the observation data from many different camera position,at last,discussion and analysis the precision of calculation according to the algorithm.The experimental results show that the calibration accuracy of the BP neural network model and improved ten parameter model is very high,and the former is slightly better than the latter,but the BP neural network model has no specific physical meaning;and the results show that the calculation accuracy based on the strict bundle adjustment with additional parameters is higher than others;Moreover,the experiments show that,the self-adaptive iteration algorithm can accelerate the convergence speed to solve the normal equation with ill-conditioned matrix,and it also enhances the calculation accuracy greatly.