In visual SLAM problems, the ORB feature has drawn much attention because of its high efficiency and stability. To address problems such as the low accuracy of image point measurements and the obvious phenomenon of feature aggregation during ORB feature extraction, a uniform distributed subpixel ORB feature extraction method suitable for high-precision SLAM was proposed. In this study, the principle of precise feature positioning was first analyzed, the error equation was then reasonably simplified, and a weight function calculation method based on template window distance was finally adopted, all of which significantly reduce the algorithm's computational cost. A quadtree-based uniform distribution solution was designed in which the image plane space is segmented with only a limited number of iterations. Features with optimal response are then exported. Experiments show that the additional computational burden of feature extraction for our method is less than 2.5 ms. The measurement accuracy of ORB features is 0.84 and 0.62 pixels on the TUM and KITTI datasets, respectively, reaching the subpixel level. Our method can thus reduce the initial value of errors and increase the efficiency of bundle adjustment. The problem of feature aggregation is effectively solved based on the condition of satisfying the overall distribution of features, which is beneficial to the robust and accurate solution of subsequent problems.