In order to improve the optical flow evaluation accuracy in special ca ses,an extension of the total generalized variation (Ext-TGV) is proposed in the optical flow analysis technique.The met hod conducts the non-local extension for the classic second-order TGV by introducing a novel high-o rder regularization term.It favors the piecewise affine solutions and allows to incorporate soft-segmentation cues into the regu larization term.The extension can accurately provide local motion boundaries and resolve the ambiguities in the matching term .In addition,the method also introduces a new matching item which is robust to illumination and scale changes.The experim ental results on the KITTI and SINTEL data show that the proposed optical flow evaluation method based on the non-loc al TGV can significantly improve the accuracy of optical flow evaluation compared with the conventional methods.Quantitatively,the average error of our proposed method is about half of that of traditional TGV method for different error thresholds on the KITTI training set,and the average end point error (EPE) of our proposed method is at least 1.5% lower than that of other outstanding methods on the SINTEL test se t.