A Novel Approach to Correction of Features from Image Matching in Unmanned Aerial Vehicle Navigation Through Clustering and Triangular Similarity

Feature extraction and matching are critical in image-based models, particularly for the navigation of unmanned aerial and ground vehicles. While classical methods have been widely used, deep learning techniques are increasingly favored due to their superior performance. However, both approaches are sensitive to perspective differences, often resulting in incorrect matches, sometimes outnumbering correct ones. Although various traditional and deep learning-based algorithms exist to mitigate this issue, they do not consistently yield optimal results. This paper presents a novel method to eliminate high-density outlier matches to enhance matching accuracy. The proposed approach leverages triangular similarities among matched features and groups randomly selected feature sets. A three-dimensional simulation environment was developed, and a dataset was prepared to evaluate the algorithm. The performance of the proposed method was compared with existing algorithms using simulation images and real-life public datasets captured from various aerial angles.