An Effective Proposal to Reliable Forward Velocity Variation of NMPC-Based Visual Path-Following Control

Vision-based control has become an interesting alternative for increasing the autonomy of mobile robot navigation in many real scenarios. Classic path-following models did not originally predict a metric for reference forward velocity variation, and this becomes an even more pronounced problem when using visual techniques that are very sensitive to parameter calibration, such as curvature. This paper proposes a novel approach to reliable forward velocity variation in NMPC (nonlinear model predictive control)-based visual path-following controllers, directly from the image plane. The main contribution arises as improvements in the image processing stage for the acquisition of practicable reference velocities and a new state capable of capturing the characteristics of the path and calculating, at runtime, an optimal forward velocity capable of safely driving the robot around the visual path. The new set of internal control inputs defined for the NMPC framework allows the application of a computationally efficient technique to handle feasibility through the relaxation of input and state constraints. Simulations and experimental results with the Husky UGV platform navigating on an imperfect visual reference path and with an arbitrary curvature profile demonstrate the correctness of the proposed method.