A Review of Statistical Modeling and Inference Methods for UAV Airborne Sensors: From State Estimation to Probabilistic Uncertainty Perception

Abstract

The autonomous flight capability of Unmanned Aerial Vehicles (UAVs) in complex dynamic environments highly depends on the accurate perception of the environment and their own state by onboard sensors. However, limited by sensor noise, model mismatch, observation heterogeneity, and external environmental disturbances, the output of onboard sensors inherently possesses significant uncertainty and randomness. How to transform imperfect, multi-source, and asynchronous sensor observations into reliable state estimation and environmental cognition results through statistical modeling and inference methods has become one of the core issues in UAV perception system research. This paper systematically reviews the development of research related to UAV onboard sensors from a statistical perspective, focusing on the application and evolution of statistical modeling, state estimation, and multi-sensor fusion methods in UAV perception systems. First, it summarizes the typical statistical observation models and noise characteristics of inertial sensors, satellite navigation, vision, lidar, and novel neuromorphic sensors, and analyzes key statistical issues such as random walk, non-Gaussian noise, and time-dependent errors. Subsequently, based on the Bayesian state estimation framework, this paper systematically reviews the application progress of Kalman filtering, error state filtering, particle filtering, and robust statistical methods in UAV navigation and localization, and compares and analyzes the statistical nature of loosely coupled and tightly coupled multi-sensor fusion strategies. Building upon this, it further discusses joint probabilistic modeling methods for heterogeneous sensors such as vision, inertial, and radar, as well as the fusion trend of statistical learning and deep models in high-dimensional perception tasks. Finally, this paper summarizes the role and limitations of statistical methods in UAV airborne sensor research and looks forward to future development directions oriented towards uncertainty perception (the capability to explicitly quantify the reliability of perception results), risk-constrained decision-making (strategies that incorporate estimation variance into control loops to ensure operational safety), and integrated sensing-computing architectures.