Exploring Probabilistic Ensembling Fusion for UAV-Based RGBT Visual Object Tracking
Da Zhang, Haibin Xie, Yangliu Kuai · 2025
In this paper, we introduce a novel RGB-T tracking framework specifically designed for unmanned aerial vehicles (UAVs), named PEECO. The proposed framework leverages a dual-stream architecture comprising two modality-specific correlation filter-based trackers, coupled through a probabilistic ensembling-based fusion module. This fusion mechanism adaptively weights the contributions of each modality by estimating their respective uncertainties. Unlike prevalent featurelevel fusion strategies reliant on deep neural networks, our method is simpler yet highly effective, eliminating the need for pre-registration between RGB and thermal modalities. Extensive evaluations on the RGBT234 benchmark demonstrate that PEECO achieves state-of-the-art performance, thus validating the effectiveness of the proposed fusion strategy in enhancing the robustness and accuracy of UAV-based RGB-T tracking.