The epipolar geometry restoration offline and online for the parallel binocular vision system were discussed in this paper. For the former, a method that uses a common calibration board to realize automatic and accurate matching of the image corner pairs was developed; for the latter, a connotative constraint was proved by geometry theory and the corresponding cross-correlation sparse matching algorithm was introduced. The constraint was that for any space point projected into two images, the horizontal coordinate in the left image of this point is bigger than that in the right image. Standard and real images experiments were made for each kind of problem and the results show that the algorithms for each problem were valid.