Abstract: Accurate and reliable dynamic obstacle perception is a crucial prerequisite for the safe navigation of autonomous mapping UAVs in confined and spatially constrained environments. However, existing LiDAR-vision fusion methods often struggle to cope with significant variations in sensor reliability under degraded conditions, such as poor illumination, reflection interference, or motion blur. To address these challenges, this paper proposed an uncertainty-modeling-driven LiDAR-vision fusion framework for dynamic obstacle detection, which adaptively adjusted sensor contributions by explicitly modeling sensor observation uncertainty. Based on probabilistic models, the framework performed real-time uncertainty quantification for both LiDAR point clouds and RGB images and introduced an adaptive sensor reliability score(ASRS) mechanism to guide fusion decisions and subsequent object tracking. Experiments were conducted on a self-constructed multi-condition dataset, and the results show that the proposed method improves the F1-score by approximately 15%-20% compared with existing methods in challenging scenarios involving low illumination, glass reflections, and motion blur. Furthermore, it maintains real-time processing performance at approximately 25 Hz on embedded platforms, validating the method's robustness and engineering feasibility in complex degraded environments.

Key words: dynamic obstacle detection, uncertainty quantification, LiDAR-vision fusion, indoor autonomous mapping UAV