Lightweight Visual Navigation with Dual-ROI Collaboration and State Machine for Autonomous Mobile Robots

Addressing the conflict between limited embedded computing power and demands for high real-time performance and robustness in visual navigation for Autonomous Mobile Robots (AMRs), this paper proposes a lightweight algorithm. Key innovations include: (1) A sliding window-based pixel-level midline extraction method adaptable to straight paths, curves, and discontinuous lines. (2) A dynamic dual-ROI collaborative perception architecture, reducing image processing by 30%–40% and enhancing the control horizon. (3) A four-state machine decision model improving sharp turn success and motion smoothness. Experiments demonstrate superior robustness in discontinuous line re-detection, complex illumination, and sharp turns compared to traditional methods, while maintaining real-time performance on the embedded platforms. This enables efficient navigation for AMRs in smart factories and warehouses.