High-Precision Feature Point Matching and Stereo-Depth Estimation Using Rotation-Invariant CNN

Stereo-matching has become essential in various industrial applications, including robotics, autonomous driving, and drone-based surveying. In the drone-based depth estimation, we captured images from two different positions and determined the corresponding points between them through stereo-matching. A longer distance between the two positions improves triangulation accuracy but makes stereo-matching difficult owing to the reduced image overlap. This limitation is inherent to previous methods, necessitating at least 50% image overlap to achieve only centimeter-level accuracy. Hence, we propose using stereo viewing with feature point matching, which allows for direct matching of points on the image. Our approach applies a novel rotation-invariant convolutional neural network (CNN) that extracts features more effectively in the presence of angular changes in a subject, surpassing the performance of previous CNN-based models. We evaluated our method using the HPatches dataset, which demonstrated an increase in feature point matching accuracy of up to 0.9%. In a practical stereo imaging setting, our method achieved a height estimation error of approximately 1.2 mm and height resolution of approximately 2.6 mm in image pairs with approximately 25% overlap under varying conditions. This performance confirms that the proposed approach effectively resolves the trade-off inherent to traditional stereo-matching techniques, particularly with regard to the challenging overlapping scenarios that these previous methods failed to account for. Consequently, this study substantially broadens the applicability and versatility of stereo-depth estimation.