To address the challenges of poor real-time computation, weak environmental adaptability, and insufficient control smoothness in the online replanning of high-speed cruising vehicles, this paper proposes an online trajectory replanning method based on model-agnostic meta-learning (MAML) warm-started sequential convex optimization (SCO). First, a three-degree-of-freedom (3-DOF) dynamic model of the vehicle is established, and a high-quality offline optimal trajectory library containing diverse obstacle avoidance scenarios is constructed using the hp-adaptive pseudospectral method. Second, a MAML training framework with inner and outer loops is constructed by incorporating nearest-neighbor residuals. By learning the dynamic manifolds and topological features inherent in the offline trajectory library, optimal common initialization parameters are obtained, enabling the network to rapidly adapt to new environments and generate high-quality initial values with only a few gradient updates. Finally, an online trajectory replanning framework based on MAML warm-starting SCO is established. Leveraging the self-learning and adaptive capabilities of MAML, the framework quickly infers warm-start initial values for SCO in new environments, enabling the SCO solver to converge within very few iterations. Simulation results demonstrate that when facing new threat environments, the proposed method effectively balances planning efficiency with control smoothness, and can dynamically plan smooth obstacle avoidance trajectories that satisfy the dynamic constraints in real time.