Siam-VO: A Siamese Network-Based Monocular Visual Odometry for High-Altitude Low-Texture Environments

In Global Navigation Satellite Systems-denied environments, high-altitude unmanned aerial vehicle (UAV) visual localization presents significant challenges, including appearance homogeneity, sparse textures, and scale ambiguity. These factors often result in insufficient feature extraction and frequent matching failures for conventional visual odometry (VO) methods, substantially degrading localization accuracy and system robustness. To address these challenges, we propose Siam-VO, a lightweight monocular VO framework tailored for high-altitude low-texture aerial navigation. Siam-VO consists of three main components: first, a Siamese network-based feature extraction and matching module that incorporates deep features with 2-D positional encoding to enhance spatial-semantic perception and employs a cross-image attention mechanism to improve matching accuracy and robustness under low-texture conditions; second, a dense, image-wide convolutional similarity matching strategy that further strengthens feature discrimination in challenging environments; and third, a multisensor fusion and sliding-window optimization module that jointly leverages pixel-level displacements from the matching network and auxiliary measurements from barometer and magnetometer sensors, enabling continuous and accurate pose estimation. We evaluate Siam-VO on the self-collected HAVIN dataset and the public FGI dataset. Experimental results show that Siam-VO achieves a mean absolute trajectory error root-mean-square error of 5.14 m on HAVIN and maintains real-time performance at 25.6 FPS on the NVIDIA Jetson Orin Nano. These results demonstrate that Siam-VO provides an accurate and efficient solution for high-altitude UAV localization in low-texture environments.