The invention relates to an abinocular stereoscopic vision matching method for generalizing belief propagation. The method comprises the following steps of: (1) acquiring a left image and a right image of two eyes to establish Markov random fields; (2) generating multiple scales of Markov random fields, wherein the kth layer is one fourth the size of the (k+1)th layer; (3) solving the Markov random fields in an order from 1 to n, decomposing the problem of searching a minimum value in a two-dimensional space into the problem of searching a minimum value in a plurality of one-dimensional spaces, reducing a state space by using a state space reduction strategy of the scale spaces and transferring a calculation result of the ith layer to the (i+1)th layer after iteration is finished; and (4)using the state with the minimum iteration value as the final state of a variable, i.e. a parallax value of points in an image, corresponding to the variable, after the Markov random field at the lowest layer is solved. The invention effectively lowers the complexity and reduces the amount of calculation.