High-accuracy multiviewpoint stereo measurement using the maximum-likelihood method

Abstract

We propose a new type of highly accurate multiviewpoint stereo measurement using the maximum-likelihood method. This method is used for measuring positions of facilities, such as telegraph poles on a road. In this method, a normal distribution is assumed for a measurement error on stereo image planes. First, the probability density functions of the measurement errors at the multiviewpoints are combined. Next, the combined function is expressed using both real-space coordinates and coordinate transformation parameters, which give relative viewpoint positions, and the most probable coordinates are estimated from a condition that minimizes the function. Then, the best viewpoint positions, which minimize the area of the combined error ellipse, are investigated by analyzing the variance included in the estimated coordinates, and the relationship between the number of viewpoints and the variance of the best viewpoint positions is obtained. Finally, these theoretical results are confirmed experimentally by measuring outdoor facilities.