Model predictive visual guidance for shortest distance

Abstract

The objective of vision-based guidance is to construct guidance command which steers the camera to the desired pose using visual measurement. This thesis concentrates on design of a guidance command which the trajectory of camera could be as straight as possible, based on a model predictive approach. A straight guidance trajectory is intended to prevent unnecessary movement of the camera. Model predictive approach is utilized due to its capability to handle system constraints, and its robustness with respect to scale uncertainty. Overall process of this thesis is 1) to generate appropriate reference feature trajectory, and 2) to construct a guidance command which follows the reference feature trajectory using an optimization technique. To generate time-independent reference feature trajectory which leads to straight guidance trajectory, partial pose estimation is performed. Estimated pose is used for feature parameterization and definition of the local reference feature is proposed. To track the reference feature trajectory, an additional term is formulated in the cost function to be optimized. Given the assumption of invariant feature depth during the prediction process, computational complexity of optimization is highly reduced, which makes the proposed algorithm more applicable to real environments. From simulation, strength of the proposed algorithm is validated. Experimental results demonstrate more direct guidance trajectory is achieved with reasonable computation time.