SIFT Feature-Based Relative Altitude Estimation Enhanced With Siamese Network

In GPS-denied environments or when GPS signals are unreliable or unavailable, alternative methods of accurate localization with coordinate generation become critical. To address localization, the scale-invariant feature transform (SIFT) algorithm, along with its numerous adaptations, is extensively utilized in computer vision and remote sensing for matching image features to identify objects and perform localization. This article presents a novel approach for estimating the relative altitude of unmanned aerial vehicles (UAVs) using SIFT features’ scale (size), omitting the need for additional data like camera intrinsic parameters, as well as extensive image datasets are also required for training. Furthermore, the approach enhances feature matching through the integration of a Siamese network, leveraging the robustness of SIFT features combined with the discriminative power of convolutional neural network (CNN) features. To further improve the performance of the Siamese network, we applied direct error-driven learning (EDL), a learning method that directly adjusts the network’s weights based on the overall error to enhance its ability to differentiate true from false matches, thereby improving the accuracy of the final altitude estimation results. Altitude estimation is achieved by comparing the SIFT features’ size in the UAV image taken from the current position with those in the preestablished reference, ensuring reliability and computational efficiency. The proposed method is computationally efficient, making it suitable for real-time applications and UAVs with limited processing capabilities. It is also highly adaptable, requiring minimal hardware and eliminating the dependency on external sensors. The robustness and effectiveness of this method were validated across four high-resolution UAV image datasets at varying altitudes.