UEDF: Uncertainty-Guided Energy-Based Dynamic Fusion for Robust Multimodal UAV Classification

Abstract

Accurate identification of uncrewed aerial vehicles (UAVs) is crucial to ensure airspace security and effective threat assessment. Traditional UAV classification systems usually rely on single-modal sensing data, which limits their ability to capture the diversity and complementarity of UAV characteristics under different environmental conditions. In contrast, multimodal sensing provides a more robust solution by integrating complementary information from heterogeneous modalities such as LiDAR and radar. However, the existing multi-modal fusion methods are usually static and cannot take into account the changes in dynamic reliability between different modalities, resulting in poor performance in complex practical scenarios. To overcome these challenges, this paper proposes an uncertainty-aware dynamic multimodal UAV classification framework that combines energy-based uncertainty quantification, history-aware difficulty tracking, and adaptive fusion. Specifically, LiDAR and radar point clouds are jointly exploited to predict UAV types, with fusion weights dynamically adjusted according to modality-specific reliability. PointNet is employed for feature extraction, while energy-based confidence estimation quantifies uncertainty to guide the adaptive fusion process. Experimental evaluations on the Multi-Modal Anti-UAV Dataset (MMAUD) demonstrate that the proposed framework achieves a mean class accuracy of 97.61%, outperforming existing single-modal and multimodal fusion approaches. Comprehensive comparisons and performance analyses further validate the effectiveness and robustness of the proposed uncertainty-guided dynamic fusion mechanism.