Adaptive Cascaded Framework for Efficient Real-Time Rotation Direction Determination of X-Type Rudders

To address the trade-off between accuracy and real-time efficiency in vision-based determination of the rotation direction of X-type rudders, this paper proposes an adaptive cascaded framework. The framework adopts an on-demand escalation strategy and integrates three analysis levels with different computational costs and levels of robustness. It first rapidly estimates the rotation direction using geometric and color cues. When confidence in the lower-level result is insufficient, the process is escalated to sparse optical-flow tracking based on the Lucas-Kanade algorithm. For more challenging scenes, dense optical flow based on Dense Inverse Search is further invoked as the final arbitration layer. In parallel, a scene context analyzer dynamically modulates the confidence score of each level according to illumination stability and texture complexity, enabling adaptive scheduling of the decision path and computational cost. Experiments are conducted on a custom-built X-type Rudder Rotation Dataset containing typical scene disturbances, including exposure variations, complex backgrounds, dynamic backgrounds, and abrupt illumination changes. The proposed framework achieves an accuracy of 90.5%, a detection rate of 90.7%, a false discovery rate of 9.4%, and a processing speed of 186.1 FPS per region of interest. Compared with single-level methods and the framework without scene-context modulation, the proposed method provides a more favorable trade-off between decision reliability and processing speed. The experimental results demonstrate the effectiveness of the proposed framework for efficient and reliable real-time visual determination of the rotation direction of X-type rudders under diverse testing conditions.