Robust feature matching for aerial visual odometry

Interest points matching for aerial visual odometry using quadrotor MAV is tackled in this work. First, a set of sparse feature points are extracted using ORB detector. These are then grouped using Gradient Vector Flow (GVF) fields by finding points of high symmetry within the image. A robust matching strategy is introduced to improve the motion estimation. In order to validate ORB features matches, their grouping points are compared. Using the matched points, windowed bundle adjustment incorporating Gauss-Newton optimization is utilised for motion estimation. In order to deal with matching outliers, a Random sample consensus outlier rejection scheme is integrated. Lack of MAV stereo datasets in the literature motivated the generation of such vital data. Detailed results validating the proposed strategy are illustrated using these datasets. Also, a comparison with other approaches is also provided and shows the superiority of our approach.