Abstract
Monitoring unmanned aerial vehicles (UAVs) at long distances presents significant safety challenges due to sensor noise and high vehicle maneuverability. This paper introduces a hybrid tracking algorithm that combines the physical interpretability of the Interacting Multiple Model Kalman Filter (IMM-KF) with the adaptive capabilities of Recurrent Neural Networks (RNNs). By utilizing a Long Short-Term Memory (LSTM) network to dynamically tune noise covariance matrices based on the filter's innovation signal, the proposed method achieves a 28.56% improvement in trajectory tracking accuracy compared to traditional static filters.
Introduction
As UAVs increasingly operate autonomously and beyond visual line of sight, reliable tracking becomes paramount. Automatic Dependent Surveillance-Broadcast (ADS-B) offers a cost-effective real-time tracking solution but suffers from significant signal noise and struggles to maintain lock during high-agility maneuvers. Traditional Kalman Filters often diverge when a UAV's motion deviates from predefined mathematical models. While the Interacting Multiple Model (IMM) approach attempts to mitigate this by switching between motion models, it typically relies on hand-tuned, static noise parameters that fail to adapt to dynamic environments.
Methodology
The proposed framework integrates an LSTM-based RNN into a standard IMM-KF loop. The IMM-KF utilizes two parallel models: a Constant Velocity (CV) model for stable flight and a Constant Jerk (CJ) model for maneuvers. The core innovation lies in the "Enhancer" component, which monitors the "innovation"—the residual error between the filter's prediction and the actual ADS-B observation.
The RNN processes this innovation signal over a sliding window to output real-time updates for the process noise ($Q$) and observation noise ($R$) matrices. This allows the system to adaptively weight the sensor data against the motion model, maintaining accuracy even during complex maneuvers where traditional filters might fail.
Experimental Setup
The algorithm was evaluated using a dataset of 10 distinct 3D UAV trajectories generated via the MATLAB UAV Toolbox. Data points were sampled at 1Hz to simulate standard ADS-B broadcast rates. The dataset was split into training (7), validation (2), and testing (1) sets, with high-precision ground truth used for supervised training of the RNN.
Results and Discussion
The RNN-enhanced IMM-KF demonstrated superior performance across all test scenarios. Quantitatively, the total Root Mean Square Error (RMSE) was reduced by 28.56% compared to a traditionally tuned IMM-KF and by 58.98% compared to raw ADS-B data.
A key observation was the system's robustness during sharp turns. While empirical filters with static parameters often diverged during high-G maneuvers, the RNN-enhanced version maintained a performance level comparable to a "perfect" filter tuned with full a priori knowledge of the flight path. This success is attributed to the RNN's ability to recognize model mismatch through the innovation signal and adjust noise parameters accordingly.
Conclusion
This research highlights the benefits of hybridizing classical control theory with deep learning. By using neural networks for parameter optimization rather than end-to-end tracking, the system retains physical interpretability while gaining significant adaptive power. Future work will focus on validating the model against real-world, non-Gaussian ADS-B noise and optimizing the computational efficiency for real-time multi-target tracking at ground stations.
Summary updated