Abstract: This paper introduces a novel optimization method for accurate 3D registration based on graphics simulation, in order to deal with non-differentiable objective functions, and meantime avoid problems of previous approaches that incur sensitivity to model rendering resolution and local extremes. Firstly we defined the objective function based on object contour, then LM（Levenberg-Marquardt） algorithm was used to solve the function. During each iteration, according to the current position and pose, the corresponding gradient is estimated adaptively, with gradient direction estimated respectively by analytical incremental estimator and random sampling method, so that the smooth expression of the objective function can be simulated. Both quantitative and qualitative experiments show that the proposed method can converge quickly and not suffer from local extremes.