To address the susceptibility of automated pointer gauge reading in port environments to complex backgrounds, strong illumination, and multi-scale variations, this paper proposes an improved YOLOv11n-Pose model. A High-level Screening Feature Pyramid Network (HS-FPN) is introduced to replace the original Neck structure, enhancing multi-scale feature fusion while reducing model parameters. Furthermore, a dual-path downsampling module (ADown) is designed to replace some convolutional layers in the Backbone, improving feature representation for small-scale targets and complex scenes. To enhance robustness to difficult samples such as occlusions and boundary cases, a hard sample mining mechanism (LRM) is incorporated into the loss function. Experimental results show that the proposed model achieves object detection and keypoint detection accuracies of 96.83% and 92.04%, respectively, on pointer gauge detection tasks in port environments. The model satisfies real-time and lightweight requirements and outperforms several mainstream detection methods, demonstrating its effectiveness for automated pointer gauge reading in port scenarios.