Considering the redundancy of localization matching between reference point and access point due to the spatial structure obstruction and weak signal penetration, a spatial dual-refinement localization algorithm named “horizontal refinement of reference points and vertical refinement of access points” is proposed. Firstly, the traditional mean value is replaced with the high-order statistical information of the strongest received signal to characterize each reference point, and the processing methods of small-area fusion and boundary reference point sharing are combined to achieve the fuzzy clustering in the target space, so as to weaken the adverse effect of absolute discrimination at the edges; secondly, based on the dimensionality-reduced subspace, the spatial differentiation and coverage reliability of each access point are measured comprehensively to screen out the subspace with high recognition value and high stability. Finally, the first-level area discrimination is performed by judging the strongest received signal source, and the second-level location estimation is achieved with the WKNN algorithm. The proposed horizontal streamlining strategy possesses the regular clustering and complies with the structural constraints of the scene more precisely according to the realistic roadshow test. It′s found that the vertical streamlining strategy improves the average positioning accuracy by at least 17% compared to the traditional access point selection algorithm, and filters out the large positioning error of more than 4. 5 m under the condition of distribution density of 1 m×1 m reference points.