The shot noise, i.e. a marked and filtered Poisson process, is particularly relevant for modelling detector signals in nuclear instrumentation. But the domain of application is much broader than particle physics.

The shot-noise very specific properties, like the absence of the analytical expression of its density, explain that very little is known about optimal estimation in this context.

The classical maximum likelyhood (ML) method and all the related results cannot be directly applied to this problem.

This thesis aims at deriving algorithms that outperform the methods presently used in nuclear instrumentation. This will be made possible by exploiting the complete shot-noise distribution instead of only taking into account the first cumulants.