Abstract: Taking the feature distribution into account, a local feature extraction algorithm is proposed. The local features satisfying isotropic distribution in the unstructured environment are extracted as natural landmarks. Thus the behavior-based robot can achieve high-precision visual homing by utilizing those landmarks. Based on the SIFT (scale invariant feature transform) algorithm, the UD-SIFT (uniform distribution-SIFT) algorithm is obtained by improving the uniformity of the feature distribution. In addition, a novel criterion for evaluating the uniformity is proposed. The visual homing experiments are carried out indoors, in the corridor and outdoors, using the ADV (average displacement vector) and ALV (average landmark vector) methods which are both based on the panoramic vision. Compared with the original SIFT, the UD-SIFT lowers the homing average angular error by more than 25.01%. The results show that this algorithm effectively improves the feature distribution and the robot homing precision.