Abstract: In unmanned transportation roadways of underground coal mines, severe dust interference, complex lighting conditions, and frequent target occlusion result in insufficient feature representation and degradation of traffic lights and anti-collision barrels in images captured by onboard cameras, thereby limiting detection accuracy. To address this problem, a real-time detection method for traffic lights and anti-collision barrels in underground coal mines based on an improved YOLOv11 (YOLOv11-CTDNet) was proposed. A new feature extraction module, SPD-NsPConv, was designed by combining Space-to-Depth Convolution (SPD-Conv) with Non-strided Partial Convolution (NsPConv) to enhance the detection performance for small objects and objects with weak features. A new feature processing module, DS-C3k2_MLCA, was developed by integrating Depthwise Separable Convolution (DsConv) with Mixed Local Channel Attention (MLCA). This module replaced the C3k2 modules in the neck network to suppress interference from background noise. Hypergraph-Based Adaptive Correlation Enhancement (HyperACE) and the Full-Pipeline Aggregation-and-Distribution (FullPAD) paradigm were introduced to improve the detection accuracy for multi-scale and partially occluded objects. A small-object detection layer was added to improve the detection performance for small objects. In addition, the conventional decoupled head was replaced with an attention-based Dynamic Head (DyHead), which collaboratively and adaptively captured critical information through three complementary attention mechanisms while ensuring adequate feature optimization, thereby enhancing the detection capability for multi-scale objects. The experimental results showed that YOLOv11-CTDNet achieved a mean Average Precision at an Intersection over Union threshold of 0.5 (mAP@0.5) of 92.4%, representing an improvement of 9.1 % over YOLOv11n. Compared with other models, such as Faster R-CNN, YOLOv11-CTDNet achieved the highest detection accuracy while maintaining a relatively low parameter count and computational complexity. Under extreme conditions, YOLOv11-CTDNet exhibited stronger adaptability to low-light and occluded scenes. The proposed model effectively improves cross-dataset generalization ability and robustness and can satisfy the practical requirements of underground auxiliary transportation systems for real-time performance, stability, and deployment under limited computational resources.