Estimating 3D Object Parameters from 2D Grey-Level Images

ii internal and external parameters is carried out with animated images to demonstrate the capabilities of the approach for this type of application. The iterative application of a linear estimator requires the computation of the Jacobian of the measurement function. The complexity of the measurement function forces a numerical approach. The numerical approach has the advantage that the elements of the Jacobian, the sensitivity coefficients, are computed model independently. The numerical approximation of the Jacobian matrix requires a choice for the size of the parameter variations. These variations tend to affect the convergence of the iterative estimator in cases where 'near dependency' exists between the columns of the Jacobian H corresponding to some parameters. This point is investigated and it is shown that careful choices of the variations may minimise the effects on the convergence of the estimated parameters. The eigenvalues of the scaled identifiability matrix, which can be computed in advance, show to have significance for the prediction of the convergence of the iterative estimation procedure. The effect on the eigenvalues of Gaussian filtering and of the parameter variation is investigated.