With the increasing depth of seismic exploration, the accuracy of middle-deep velocity modeling is low due to the complex geological structure and drastic lateral velocity variation in the subsurface, which affects the quality of migration imaging. At present, among the depth domain velocity inversion methods, ray travel time tomography has high computational efficiency, but it is prone to problems such as shadow region and caustics. However, the full waveform inversion has high accuracy, but it has some shortcomings such as large calculation. Therefore, a velocity modeling method that takes into account the computational efficiency of ray tomography and the accuracy of full waveform inversion is needed. Gaussian beam tomography is an intermediate method between ray theory and wave theory, which improves the stability of tomography method by using kernel functions to reduce the pathologies of the sensitivity matrix. In this paper, on the basis of using Gaussian beam to extract the Angle domain common imaging gather, a data-driven seismic DNA algorithm is used to automatically fit the residual curvature of the gather and extract the depth residual, and the depth residual is converted into the travel time residual in the inversion equation, and then the cumulative error carried by the downward propagation of seismic waves is eliminated by a reverse recursion method. Further improve the accuracy of the migration velocity. The accuracy, effectiveness and adaptability of the method are verified by the model and practical data.