Fatigue driving detection demands high real-time performance and reliability, yet existing object detection models suffer from large parameter counts and high computational complexity, making them difficult to deploy on in-vehicle edge devices. To tackle this problem, this paper proposes Flat-YOLO, an ultra-lightweight detection model tailored for in-vehicle environments, which reduces parameter redundancy through progressive flattened channel scaling and incorporates two complementary modules, bounding box-aware channel self-attention (BBoxCSA) and IoU-aware multi-scale fusion (IoUMSF), to enhance edge representation and multi-scale fusion. On the self-constructed driver drowsiness detection (DDD) dataset (9 751 images, 7 classes), Flat-YOLO reduces parameters by 69.4% compared to YOLOv8n while maintaining comparable mAP@0.5:0.95, with precision and recall improvements of 0.3% and 1.0%, respectively; compared to YOLOv8s, it achieves a 91.7% parameter reduction and 76.8% computational cost reduction. Class-wise results further demonstrate that for safety-critical categories, Flat-YOLO improves recall for head_low by 4.9% (0.951 vs 0.902) and for open_eyes by 2.2% (0.913 vs 0.891); for fine-grained fatigue cues, it enhances mouth′s mAP@0.5:0.95 by 0.9% (0.479 vs 0.470) and yawn′s precision by 3.3% (0.947 vs 0.914). Furthermore, on the public yawning detection dataset （YawDD） dataset, Flat-YOLO achieves an mAP@0.5:0.95 of 0.901 (compared to 0.862 for YOLOv8n), validating the cross-dataset generalization capability of the proposed method.