Error characterization of flight trajectories reconstructed using Structure from Motion

Abstract

This research assesses the accuracy of Structure from Motion algorithms in replicating aircraft flight trajectories in real world coordinate systems. Structure from Motion techniques can be used to estimate aircraft trajectory and attitude by estimating the position and pose of a camera mounted on the airframe from a series of images taken with that camera. The scale and coordinate systems associated with these pose estimates are arbitrary but can be tied to a real world coordinate system and scale with knowledge of terrain features or aircraft INU/GPS measurements. As a result, Structure from Motion techniques hold great promise for use in image based Simultaneous Localization and Mapping (SLAM); however, the error associated with these techniques must be understood for incorporation into a robust navigation system. The error in the Structure from Motion trajectory reconstruction is affected by feature matching methods, camera geometry and algorithm parameters as well as the method used to convert from reconstruction coordinates to real world coordinates. Methods of using inertial and geographic measurement data to associate a scale and real world coordinate system with the trajectory estimates are developed and compared. The reconstructed flight trajectory output by Structure from Motion is compared to the actual flight trajectory to characterize errors using both simulation and flight test data in a variety of flight regimes.