Next: 3 Problem Definition Up: Improving model shape acquisition Previous: 1 Introduction

2 Background

  The integration of geometric constraints into the shape fitting process has been treated for wire frame model construction by Porrill [ 8 ]. Wire frame models were constructed from stereo-data. The model features were given statistical distributions and geometric constraints between features produced dependencies in the distributions. The model adjustment process maximized the a posteriori probability of the models. Since the models were based on wire frames, the constraints were related to lines. Four types of constraints were considered: orthogonality, intersection, equality and connection by a small rigid motion. The optimal feature parameters were estimated using an extended Kalman filter. At each iteration, constraints are linearized in the neighborhood of the current estimate, and then used to correct the measurement. Porrill's approach is nice since it takes advantage of the recursive linear estimation of Kalman filtering, however it assumes a Gaussian distribution which may not always the case. Moreover, the method, guarantees the satisfaction of the constraints only to linearized first order. Additional iterations at each estimation step are needed if one would like to obtain more accuracy. This last condition has been taken into account in the work of De Geeter and al [ 4 ] by defining a ``Smoothly Constrained Kalman Filter''. The key of their approach is to replace a nonlinear constraint by a set of linear constraints applied iteratively and updated by new measurements in order to reduce the linearization error.



Naoufel Werghi
Thu Jul 10 17:36:04 BST 1997