Precise localization is an important premise of autonomous navigation for agricultural vehicles. We propose a localization algorithm based on monocular camera. After features are detected and tracked through multiple frames, vehicle poses are estimated based on 3D-2D correspondences. Furthermore, the translation absolute scale is calculated based on the assumption that the ground patches are locally flat and the camera is moving at a known and fixed height over the ground. Finally, the poses are refined by the pose threshold filter. Compared with the RTK-GPS data, the average relative position errors of the three different experimental terrains are 5.459 9%, 8.373 1% and 6.443 94%, and the average heading errors are 7.717 7°, 5.738 9° and 3.438 3°. The results show that the algorithm is feasible for agricultural vehicles localization.