Scientific Machine Learning-Supported Heterogeneous Track-to-Track Fusion Using Radar and Infrared Sensors

In modern surveillance and defense systems, accurate tracking of aerial and ground targets remains a critical challenge due to the presence of uncertainties and dynamic target maneuvers. This study presents a robust air-to-air and air-to-ground tracking framework that integrates heterogeneous sensor measurements from radar and infrared sensors. The core methodology is based on an Extended Kalman Filter with a constant velocity motion model, further enhanced by heterogeneous track-to-track fusion. To mitigate estimation errors arising from dynamic uncertainties, a scientific machine learning-supported approach is proposed to tune the process noise covariance and sensor trackers' covariances adaptively. The proposed method refines the covariance coefficients to improve tracking accuracy under various maneuvering conditions, including climbing, descending, constant rate turns, and accelerated motions. The optimization is conducted over multiple simulated scenarios, where optimal parameters and estimated state vectors of the tracked objects serve as input features. Neural network models are then trained on these features to generalize the optimization results, enabling real-time estimation of process noise covariance coefficients in unseen scenarios. Experimental evaluations demonstrate the effectiveness of the proposed approach in adapting to dynamic target maneuvers, reducing estimation errors, and improving overall tracking performance.