Abstract： Aimed at the problem of navigation of unmanned aerial vehicle(UAV) in a complex unknown environment, an algorithm of position and attitude estimation based on binocular vision was described and improved in this paper. The feature points in the stereo image pairs were detected and described using the KAZE features in the nonlinear scale space. The feature points were matched with the Knn algorithm. The 3D stereo information of the feature points was calculated in the camera coordinate system. The position and attitude of UAV were estimated with the RANSAC algorithm and the LM iteration algorithm. Some experiments were conducted. The result shows that KAZE features have better accuracy, realtime and repeatability than those of SIFT and SURF. The improved algorithm can meet the requirements of UAV realtime navigation.
Key words：