<p>There are many challenges, such as low light, high noise, and complex obstacle distributions, when driverless electric locomotives operate in complex underground coal mine environments. These factors often result in low detection accuracy, especially for multiple or small targets. This paper proposes an improved real-time obstacle detection model, RSAE-YOLOv11n. First, the model is optimized by integrating an RCM into the C3k2 structure, this facilitates the capture of a broader spectrum of contextual information, enhancing the representation of intricate features and improving the extraction of multi-scale characteristics of the target. Second, the SPPF module is refined with SENetv2 to enhance the feature extraction performance of minor target detail features. Third, the ADown module substitutes conventional convolution to realize the model’s lightweight further. Finally, the EMA is implemented to improve the generalization and robustness of the model effectively. Experimental findings demonstrate that RSAE-YOLOv11n achieves a mAP of 90.7% on a self-constructed coal mine obstacle data set, surpassing the original YOLOv11n by 2%, with a 4.2% in P, a 1.2% in R, a 12% reduction in the parameter count, a 14.3% decrease in GFLOPs, an 11% diminution in model size, and the FPS reaches 85.7. Compared to other target detection algorithms, such as YOLOv8 and YOLOv10, RSAE-YOLOv11n exhibits significant advantages in detection accuracy, detection speed, computational efficiency, and model lightweight, providing a strong theoretical foundation for the intelligent perception of driverless electric locomotives in underground coal mines.</p>