There are weak or no GPS signals indoors, so the accuracy of indoor UAV positioning is low.To solve the problem, the fusion of visual and inertial data is introduced to realize indoor UAV positioning.At the front end, the feature matching algorithm is improved.As for the rotation motion of the image, the Principal Component Analysis (PCA) method is used to calculate the rotation angle, the grid of the feature point and its 8 neighborhood grids are changed, and the Gaussian threshold is set according to the Euclidean distance between the neighborhood feature point and the matching feature point.A new score statistics model is proposed to increase the number of correct matching pairs, so as to improve the rapidity of feature matching and the accuracy of indoor visual positioning.To solve the problem of mismatching caused by local similarity of images, a data set is determined by using the geometric relationship between feature points.The similarity of data is analyzed by using Pearson correlation coefficient, and the threshold is set to eliminate the feature matching pairs with low confidence, so as to optimize the visual estimation of UAV pose information.At the back end, the visual inertia is used to optimize the pose information based on the tight coupling of the sliding window.The UAV hovering experiments under normal indoor illumination and dim indoor illumination are designed, and the flight log is analyzed.It can be seen that the feature matching speed of the improved Grid-based Motion Statistics (GMS) algorithm is 3 times higher than that of the original algorithm, and the mismatching in similar local areas is eliminated.The feature matching accuracy can reach 94%, and the accuracy of indoor UAV positioning can reach 0.02 m.The algorithm can be better applied in complex military environments.