The unmanned aerial vehicle needs an effective visual odometry system intask of indoor. The sparse visual odometry of RGB-D sensor is difficult to extractor enough color image features in darkness environment of indoor, which results in the failure of matching. We present a method that extract 2D features from the depth image directly. The paper discussed the matching performance of ORB, FREAK and SURF descriptor to the depth image. The results of the experiments show that it is feasible extractor 2D features from the depth image to matching in visual odometry, the ORB descriptor is better than other methods suitable for this kind of application.