Abstract: Foreign object detection on coal mine conveyor belts is affected by underground environmental factors such as coal dust occlusion, uneven illumination, and conveyor belt vibration, making it difficult to stably extract multi-scale foreign object features while simultaneously achieving high detection accuracy and real-time performance. To address these problems, an improved YOLOv11n-based model, termed ADCG-YOLO, was proposed for foreign object detection on coal mine conveyor belts. The convolution modules in the backbone network of YOLOv11n were replaced with adaptive kernel fusion convolutions to enhance multi-scale feature extraction through dynamically adaptive convolution kernels. A dynamic multi-branch hybrid module was introduced into the C3k2 module of the backbone network to improve the representation capability for foreign objects with diverse morphologies by employing multiple parallel depthwise convolution branches. A channel mixing module was inserted between the backbone and neck networks to achieve multi-scale feature calibration and channel enhancement. In the neck network, an aggregation-distribution dynamic attention scale sequence fusion structure was adopted to strengthen cross-scale feature correlation for small targets and improve the utilization of multi-path features. Experimental results showed that the ADCG-YOLO model achieved an mAP@0.5 of 92.56%, an mAP@0.5:0.95 of 58.72%, a precision of 89.24%, and a recall of 82.44%, representing improvements of 4.17%, 2.14%, 2.67%, and 4.44%, respectively, over YOLOv11n. In addition, the model achieved a frame rate of 136 frames/s, demonstrating a favorable balance between detection accuracy and real-time performance. In scenarios involving small elongated targets, multi-scale occluded targets, low-saliency targets, low-contrast targets, and targets with irregular shapes, the ADCG-YOLO model effectively reduced missed detections and false detections while providing high overlap between the predicted bounding boxes and the target regions. Grad-CAM++ heatmap analysis further showed that the ADCG-YOLO model enhanced its ability to focus on target regions while effectively suppressing irrelevant responses from background regions.