Accurate trajectory recovery is crucial for location-based services like traffic prediction and route optimization. However, raw trajectory data contain missing or noisy positions due to environmental interference and sensor limitations. Existing methods perform poorly in complex urban environments, such as irregular overpasses and parallel roads with varying movement patterns. To overcome these limitations, we propose MSTRLG, a Multi-Scale Trajectory Recovery via Local-Global Similarity Fusion framework. Our solution introduces (i) an adaptive multi-scale grid representation that dynamically weights optimal spatial scales to capture geometric features in overpass trajectories, and (ii) a novel fusion mechanism that combines local geometric patterns from similar trajectories (identified by learnable Fourier features) with global motion continuity constraints. Experiments on real-world datasets show that MSTRLG outperforms state-of-the-art methods, reducing RMSE by \(16.01\% \) for trajectories with ten consecutive missing points.