Multi-image Matching Using Neural Networks and Photogrammetric Conditions

Automatic determination of three dimensional information from digital images is a fundamental problem in digital photogrammetry and computer vision. The hardest part of the problem is finding conjugate points in two images. Despite the wealth of information contained in digital images, factors such as occlusion and discontinuity weaken several matching algorithms. However, image matching using more than a pair of stereo images enhance the reliability of the image matching process. This paper presents an alternative approach to match image points across several views. For each pair of images, the coplanarity condition and the correlation coefficient of image intensities are computed for each pair of image points. These two measures are feed into a feedforward neural network used to solve the multi-image correspondent. The collinearity condition is then used to validate the outputs of the neural network and to compute the 3D coordinates of the matched points. The detection rate of the neural network is about 95% to 98% and the false alarm rate is about 7% to 4%. In addition, the collinearity condition eliminated several of the incorrect matches and reduced the false alarm rate to less than 2%. The RMS errors of the ground coordinates are seven to eight centimetres.