To address the challenges of traditional physiological signal monitoring—such as its invasive nature—and the limitations of existing computer vision solutions, including unreliable extraction of micro-expression features (e.g., eyelid closure, yawning frequency), high computational redundancy, and the resource constraints of edge devices, this paper proposes a real-time fatigue detection system based on an improved YOLOv11 algorithm, designed for operational scenarios such as cockpits. The core contribution involves an innovative redesign of the backbone network through the introduction of the EfficientNet compound scaling mechanism, which optimizes the network across depth, width, and resolution. This enhancement significantly improves feature detection capability for small-scale targets. Additionally, a new SDLoss function is designed to increase localization accuracy, effectively alleviating the high sensitivity of IoU to positional deviations in small objects. Experimental results demonstrate that the improved model achieves a 1.619% increase in mAP under challenging conditions including occlusion and varying illumination, while maintaining inference speed within millisecond-level response requirements. A synchronized embedded real-time monitoring system was also developed, incorporating an interactive multi-section interface with audio-visual alerts to provide closed-loop fatigue warnings in ship piloting scenarios, substantially enhancing the effectiveness of safety supervision.