Adaptive multi-filter EKF framework for UAV tracking of ground robots in challenging terrains

Accurate relative pose estimation between uncrewed aerial vehicles (UAVs) and uncrewed ground vehicles (UGVs) in off-road, Global Positioning System (GPS)-denied settings remains difficult due to rough terrain, wheel slip, and sensor noise. This article presents a multi-filter framework built on three coupled Adaptive Extended Kalman Filters (AEKFs). The main contribution lies in a triple-layer adaptation design. First, the UAV process noise covariance updates online using short-term variance from inertial measurement unit (IMU) signals. Second, UGV motion estimation includes a terrain slip compensation term derived from wheel angular velocity and estimated ground inclination. Third, the relative pose filter adjusts its process noise scale based on the strength of UAV relative maneuvers. The experimental results show a clear gain over fixed-covariance filters with mean three-dimensional Root Mean Square Error (RMSE) reduced by 43.8 percent, decreasing from 0.32 to 0.18 m. A detailed comparison with a state-of-the-art Particle Filter (PF) further highlights the benefit of the proposed approach. Relative pose accuracy improved by 42.5 percent, where the AEKF achieved an RMSE of 0.7623, while the PF reached 1.3256. The Computational efficiency was also assessed. The AEKF reduced average CPU load by 87.14 percent, operating at 12.83 percent compared to 99.97 percent for the PF. Execution time per update decreased from 0.047652 to 0.000043 s, enabling high-rate estimation without resource saturation. Comparing with the state-of-the-art adaptive EKFs, the proposed adaptive EKF achieved a total RMSE of 0.762 m, improving accuracy by 33.5 percent, 31.2 percent, 9.8 percent, and 31.5 percent over innovation-based, multi-modal, covariance-inflated, and NN-assisted EKFs. These results confirm that the proposed framework delivers accurate, stable, and efficient relative pose estimation. The proposed method is also simulated in ROS-2 environment using Ardupilot Gazebo plugin with mission planner ground control station (GCS). The proposed approach is well suited for real-time UAV–UGV cooperation in GPS-denied tasks, including autonomous landing, search and rescue, and off-road mapping. This balance supports deployment in long-duration missions with strict timing and power limits.