<p>In open-pit mining operations, the ocular-state cues of mining-truck drivers constitute a critical indicator for fatigue assessment and are tightly coupled with operational safety. However, the eye regions captured in in-cab imagery are often low-resolution and severely blurred, which substantially hinders reliable ocular-state estimation. To address this challenge, we propose EGCM-SegFormer, an edge-guided cross-scale semantic segmentation approach tailored to low-resolution driver-eye images. Specifically, we devise a plug-and-play Edge-Guided Cross-Scale Modulation (EGCM) module, which strengthens structural consistency and feature discriminability via edge-prior-constrained bidirectional multi-scale feature fusion and gated modulation, thereby alleviating the degradation of fine-grained structures. Moreover, we introduce a mixed supervision scheme that combines cross-entropy, Dice, and IoU losses, improving optimization stability under blurred boundaries and class-imbalanced conditions while maintaining pixel-level discrimination. To rigorously evaluate the proposed method, we construct MiningEye-LR, a low-resolution semantic segmentation dataset for mining-truck driver eye regions. Extensive experiments demonstrate that EGCM-SegFormer achieves clear improvements over baseline models in both accuracy and efficiency, and cross-domain evaluations on FIVES and LoveDA further verify its strong generalization for fine-grained structural modeling and complex boundary delineation. Overall, EGCM-SegFormer provides robust structure-aware perception in challenging low-resolution scenarios, offering an effective fine-grained visual sensing solution for fatigue-driving monitoring in industrial environments.</p>