Model Predictive Visual Servoing for Dynamic Targets with Quadrotor UAVs

This paper presents an integrated framework for model predictive visual servoing of dynamic targets using a quadrotor unmanned aerial vehicle (UAV). The proposed architecture integrates visual perception, nonlinear state estimation, and constrained model predictive control (MPC) to achieve accurate and smooth real-time target tracking. In the perception module, an ArUco marker attached to the target is detected by an onboard monocular camera, and the relative pose is estimated via the solvePnP algorithm. To mitigate measurement noise and obtain reliable velocity information, an unscented Kalman filter (UKF) is employed to fuse noisy pose observations and predict the short-term motion of the dynamic target. Based on the predicted trajectory and the current state, a linear MPC controller generates acceleration commands that minimize visual tracking errors while explicitly satisfying velocity and acceleration constraints. The computed commands are transmitted to the PX4 flight controller for low-level attitude stabilization. The complete framework is implemented in ROS 2 and validated through numerical simulations in the Gazebo environment. Comparative results demonstrate that the proposed MPC-based visual servoing approach achieves improved tracking accuracy and smoother control behavior compared to a conventional PID-based method.