Robust Target Perception and Tracking for Quadrotor UAV in Dynamic Environments

In order to solve the core challenges of inaccurate perception and control lag of dynamic target tracking in complex environments, this paper proposes an autonomous tracking framework for UAV with deep coupling of visual perception and control. At the perception layer, the extended Kalman filter (EKF) based on the residual adaptive mechanism is used to accurately predict the target maneuver, and the re-identification (Re-ID) network embedded with the attention module is used to effectively deal with occlusion. At the control layer, binocular vision is used to compute speed feedforward to compensate for system delay, and Jerk limits are introduced to ensure smooth flight. Experimental results show that the proposed perception algorithm shows comprehensive better performance than the existing mainstream methods. In the dynamic tracking test, the proposed control scheme reduces the root mean square (RMSE) of the tracking error from 7.66 pixels to 4.42 pixels, which greatly improves the tracking accuracy and smoothness.