Homography-Driven Plane Feature Matching and Pose Estimation

The significance of feature matching is self-evident in the tasks of applying Simultaneous Localization and Mapping (SLAM) technology. However, existing methods mainly focus on nonplanar features and do not deal well with the matching of plane features. To elegantly handle this situation, we introduce a new constraint based on the homography matrix, called symmetric transfer error. The restriction is added to a feature matching model to form a new model named homography-driven classification network(HDCN). The model matches plane features by finding correspondence and eliminating outliers. Because of the particularity of the plane feature, we make an indoor plane dataset to train this model effectively, which consists of a large number of text labels. Through end-to-end training, the inliers ratio and the accuracy of camera pose are greatly improved. Our approach far exceeds other methods of learning and traditional algorithms in the pose estimation task of the indoor environment.