Abstract: To address the core challenges in coal conveyor belt foreign object detection, including difficulties in small target recognition, poor adaptability to low-illumination environments, high missed detection rates in occlusion scenarios, and the trade-off between model real-time performance and accuracy, a detection algorithm based on LDAFL-YOLOv11 (LSKA and DYT-enhanced ASF with FFD and LSDECD YOLOv11) is proposed. Systematic enhancements to YOLOv11 are implemented from four aspects: feature extraction, feature fusion, detection head optimization, and lightweight adaptation. Specifically, the SPPF_LSKA (SPPF with Large Separable Kernel Attention) and C2PSA_DYT (C2PSA with DynamicTanh) modules are embedded into the backbone network to enhance the capability of capturing foreign object features in occluded and low-light scenarios. The ASF (Attentional Scale Sequence Fusion) mechanism is designed in the Neck layer to address the alignment and weight allocation of multi-scale foreign object features. The C3k2_FFD (C3k2 with Frequency-Focused Dynamic Convolution) module is constructed, integrating frequency dynamic convolution to achieve differentiated learning of high-frequency details from foreign object edges and low-frequency textures from the background, while maintaining lightweight advantages. The LSDECD_Head (Lightweight Shared Detail-Enhanced Convolutional Detection Head) detection head is redesigned to reduce computational redundancy and improve small target detection accuracy through shared convolution and detail-enhanced convolution. Experimental results demonstrate that the improved model achieves 91.1% mAP50 and 66.7% mAP50–95 on the constructed hybrid dataset, with only 2.33 M parameters and an inference speed of 172 frame/s. Compared to the baseline YOLOv11n, the precision is improved by 4.0%, mAP50 by 2.7%, and the parameter count is reduced by 9.69%. When deployed on an Orange Pi 5 Plus edge device, the model maintains stable performance at over 30 frame/s. In comparison with mainstream lightweight models such as YOLOv8n and YOLOv10n, as well as several improved algorithms tailored for coal mine scenarios, the proposed method demonstrates superior performance in detection accuracy, lightweight design, and real-time capability. It effectively addresses the challenge of foreign object detection within the complex environment of coal mine conveyor belts.