Page 84 - Classification Parameter Estimation & State Estimation An Engg Approach Using MATLAB
P. 84

DATA FITTING                                                  73


                                                              2ε
                                ε 2
                                                               ρ (ε)
                                   ρ(ε)                        ε



                                                                       ε








                         –σ   +σ        ε               –σ    +σ

            Figure 3.9  A robust error norm and its derivative

            derivative   e (e) ¼ q e kk  =qx of  (:) is nearly a constant. But for large
                                robust
            values of e, i.e. for outliers, it becomes nearly zero. Therefore, in a
            Gauss–Newton style of optimization, the Jacobian matrix is virtually
            zero for outliers. Only residuals that are about as large as   or smaller
            than that play a role.

              Example 3.9   Robust estimation of the diameter of a blood vessel
              If in example 3.8 the diameter must be estimated near the bifurcation
              (as indicated in Figure 3.8(a) by the white arrows) a large modelling
              error occurs because of the branching vessel. See the cross-section in
              Figure 3.10(a). These modelling errors are large compared to the
              noise and they should be considered as outliers. However, Figure
                                                        2
                                                      e
              3.10(b) shows that the error landscape kk has its minimum at
                                                        2
               ^
              D D LS ¼ 0:50 mm. The true value is D ¼ 0:40 mm. Furthermore, the
              minimum is less pronounced than the one in Figure 3.8(c), and there-
              fore also less stable.
                Note also that in Figure 3.10(a) the position found by the LS estimator
              is in the middle between the two true positions of the two vessels.
                Figure 3.11 shows the improvements that are obtained by applying
              a robust error norm. The threshold   is selected just above the noise
              level. For this setting, the error landscape clearly shows two pro-
              nounced minima corresponding to the two blood vessels. The global
                                     ^
              minimum is reached at D robust ¼ 0:44 mm. The estimated position
                                     D
              now clearly corresponds to one of the two blood vessels as shown in
              Figure 3.11(a).
   79   80   81   82   83   84   85   86   87   88   89