To navigate autonomously, an Unmanned Aerial Vehicle, also known as drone, must be able to estimate its position during flight. This is usually achieved with the use of the global positioning system combined with an inertial measurement unit. Although widely used for this purpose, the GPS has some disadvantages, like the presence of GPS-denied cluttered environments. A way to circumvent this issue is to employ visual odometry techniques with feature-based methods to compute the aircraft motion and thereby allow the position estimation. We employ two recent features detectors and descriptors - ORB and AKAZE - and compare their accuracy and processing time. The objective is to evaluate their suitability for visual odometry task in real-time embedded systems of UAVs. We have concluded that, although ORB is faster to compute, AKAZE shows a better compromise between speed and performance than ORB for images with low resolution.

ORB, AKAZE, UAV, visual odometry, feature matching.