A new shape descriptor obtained by skeletonisation of noisy binary images i
s presented. Skeleton extraction is performed by using an algorithm based o
n a new class of parametrised binary morphological operators, taking into a
ccount statistical aspects. Parameters are adaptively selected during the s
uccessive iterations of the skeletonisation operation to regulate the chara
cteristics of the shape descriptor. A probabilistic interpretation of the s
cheduling strategy used for parameters is proposed by analogy to stochastic
optimisation techniques. Skeletonisation results on patterns extracted by
a change-detection method in a visual-based surveillance application are re
ported. Results show the greater robustness of the proposed method as compa
red with other morphological approaches.