In the automatic target recognition with complex background, the method of detecting motion target using the global
motion estimation (GME) in image sequences is often proposed. Due to the possible presences of differently moving
foreground objects and other sources of distortions, improvement of robustness and preciseness of GME is very difficult.
Therefore the method of GME based on the M-estimators' formulation with direct multi-resolution is proposed in our
paper. The M-estimators' formulation is not only executed in gradient descent at each level of the pyramid, but also
applied in initial translation estimation with minimal SSD at the coarsest level, which assures the convergence of the
subsequent gradient descent algorithm. Comparative experiments are performed to validate the performance of the
proposed algorithm. The effectiveness and improvements can be observed from the comparisons.