SiamDUKF: RGBE Siamese tracking with a differentiable unscented Kalman filter

In practical UAV tracking, performance is often degraded by nonlinear maneuvers, abrupt illumination changes, and occlusion. Single-modality trackers therefore struggle to maintain robust performance in complex environments. To address this issue, we propose SiamDUKF, a Siamese-network-based tracking framework that integrates RGB and event information with a differentiable unscented Kalman filter (Diff-UKF). Built upon SiamFC, the framework adopts a dual-branch backbone with unshared weights to extract modality-specific features and employs CBAM for adaptive cross-modal refinement and fusion. In addition, a lightweight MLP predicts the process and measurement noise covariance based on tracking confidence and motion nonlinearity, enabling end-to-end joint optimization of multimodal feature learning and state estimation. Experimental results on the VETD220 dataset show that SiamDUKF achieves a success rate of 62.2%, a precision rate of 83.0%, and a tracking speed of 82.7 FPS, indicating competitive accuracy, robustness, and real-time performance.