Credibility State Estimation Fusion Driven by Deep Learning for Low-Altitude Target Tracking

Abstract

This paper presents a novel fusion estimation framework that integrates credibility-based Kalman filtering with Long Short-Term Memory (LSTM) networks to address uncertain dynamic model parameters in low-altitude target tracking. First, the iterative formulation of a Cubature Kalman Filter (CKF) is reformulated to enable real-time adaptation to dynamic uncertainties alongside simultaneous system state estimation. Second, an estimation framework utilizing a dropout-based LSTM network is developed to improve the accuracy and generalization capacity of the estimator. Finally, a unified credibility metric is established to evaluate the estimates from both methods under a common criterion. By employing a sliding-window technique on real-time measurement data, this metric dynamically dictates the fusion weights, thereby achieving an adaptive and robust integration of model-driven and data-driven approaches. Note to Practitioners—Low-altitude target tracking in complex environments faces a core challenge: target motion dynamics are rarely deterministic. Wind, maneuvers, and sensor imperfections create uncertain dynamics that idealized mathematical models cannot capture. While traditional tracking filters (like Kalman filters) lose accuracy when the model is wrong; purely data-driven methods fail when training data does not cover real-world conditions. This work presents a fusion framework that integrates the strengths of both paradigms. A credibility-based Kalman filter continuously assesses the reliability of its internal model, while a Long Short-Term Memory (LSTM) network learns to predict the optimal filter gain directly from measurement history. The two estimates are then fused in real time using their respective credibility scores - with the more credible estimator receiving a larger weight. This credibility is assessed using a sliding window over recent measurements. Key practical insights: • No perfect model is required; the framework works with approximate dynamics. • LSTM outputs only the Kalman gain-not the full state-preserving interpretability and physical consistency. • Unified credibility metric enables transparent, adaptive fusion. • Validated on Gaussian, non-Gaussian, bimodal uncertainties, and real fixed-wing UAV flight data. Limitations and future extensions: • Needs extension to sensor uncertainty. • LSTM requires offline pre-training; periodic retuning may be needed for unseen conditions. • Sliding window introduces slight latency; window length tunable for fast dynamics. Broader applications: Robotic manipulation, autonomous vehicles, power systems-any domain with imperfect physical models and available sensor data. Bottom line for practitioners: One need not choose between model-based rigor and data-driven flexibility. By letting each estimator earn its weight in real time, this framework delivers accurate and trustworthy state estimates even when the physics are messy and the data are imperfect.