Goal of seismic data processing is to convert a recorded wave field into a structural or lithological image of the subsurface. This requires a model of the wave propagation velocities of the subsurface. Nevertheless obtaining this model is often the most difficult processing step, in areas of complex structure such as foothill or salt body. The goal of this thesis was to develop a robust and efficient migration velocity estimation technique that uses global optimisation methods and remains computationally tractable. We used e-MOEA (MultiObjective Evolutionary Algorithm) as an optimisation tool. We attempted to reduce the computational cost of EA by adding geological and geophysical knowledge to the component of the algorithms, at the representation level, in the variation operators and in the objective function. In this work, after some effort to design a concise and geologically meaningful representation, we finally concluded that the grid representation was the most flexible one, even though it implies a large number of unknown parameters and induces a high computational cost. We modified the differential semblance function and used it together with the semblance function to have a robust and accurate criterion. We hybridized e-MOEA using RMO (Residual MoveOut) and dip information. We first customize the e-MOEA algorithm itself, and also proposed a new exploitation operator and technique of optimal parent synthesis. The goal of the customization was to strive to have both the robustness of global methods and the efficiency of local optimisation methods. We presented examples of migration velocity analyses mainly on Marmousi model, together with a few results on the North-Sea L7 model. We demonstrated that our automatic velocity analysis technique is able to cope with large velocity errors and, in term of computation cost, it is as efficient as the gradient methods except in salt body. We expect that with some clever adoption this algorithm can be extended for 3D.