On the efficiency of a misspecified contaminated nonlinear regression model : application to time-delay and doppler estimation