Abstract: Aiming at the problem that real-time performance and detection accuracy are difficult to be bal-anced in multi-scale foreign body detection in coal mine transportation systems, and small-target foreign bodies are prone to missed detection and false detection, this paper proposes a foreign body detection method ADCG-YOLO based on YOLOv11n. Taking YOLO11n as the baseline, this method constructs a four-stage architecture of "backbone extraction - transition enhancement - feature fusion - detection identification": an Adaptive Kernel Fusion Convolution (AKFConv) is proposed to improve feature adaptability; a Dynamic Inception Mixer Module (DIMM) is con-structed to enhance the feature extraction capability of the backbone network; a Channel Mixer Module (CMM) is introduced as a transition enhancer between the backbone network and the neck to realize preliminary calibration and channel enhancement of multi-scale features; finally, a feature fusion network is designed in the neck, which integrates the Gather-and-Distribute (GD) mechanism and the Dynamic Attentional Scale Sequence Fusion (DASSF), so as to strengthen the cross-scale feature correlation of small-target foreign bodies and improve the utilization rate of multi-path features. Experimental results show that compared with the baseline YOLOv11n, the mAP@50 of ADCG-YOLO is increased from 88.39 to 92.56, the mAP@50:95 is increased from 56.58 to 58.72，and FPS increased from 103 to 136.. It indicates that the proposed method achieves a good balance between detection accuracy and real-time performance, and can provide algorithmic support for online monitoring and edge deployment of foreign bodies on under-ground coal mine conveyor belts.