To mitigate the problems that perception delay is readily amplified into spatial deviation, fixed regions cannot adapt to lane projection variation, and single-path dynamic updating cannot simultaneously ensure temporal continuity and transient responsiveness on ultra-high-speed curved highway sections, a dual-path fusion-based adaptive region-of-interest method is proposed for the perception layer of lane-keeping systems by constructing a pose-aware dynamic projection model, a cooperative stable-fast path updating scheme, and an uncertainty-driven fusion mechanism. The results show that, under standard operating conditions, the proposed method achieves an ROI coverage ratio, background redundancy ratio, and temporal jitter of 0.994, 0.066, and 0.0058, respectively; its root mean square lateral deviation is 0.071 m, representing reductions of 66.2% and 52.1% compared with Fixed-ROI and No-ROI, respectively; under illumination disturbance and enhanced attitude disturbance, the degradation rates of lateral deviation are only 10.1% and 14.0%, while the lane departure rate remains 0 and the task completion rate remains 100% across all standard conditions. These results indicate that the proposed method can achieve coordinated improvements in region completeness, update smoothness, and closed-loop lane-keeping performance on high-speed curved highway sections.