Aiming at the difficulties of high computational complexity, low frame rate, and reduced accuracy of existing coal mine conveyor belt foreign object detection models after lightweighting, this study proposes a new and efficient foreign object detection algorithm, HFD-YOLO, on the basis of the improved YOLOv8n. The method replaces the original backbone network with the lighter HGNetv2, designs the lightweight C2f_Star module, and uses shared convolution to redesign the detection header. BiFPN is introduced and the detection head is redesigned using shared convolution. In addition, an adaptive magnitude pruning method was used for model pruning followed by fine-tuning. The experimental results show that compared with the YOLOv8n model, the mAP50% of HFD-YOLO is improved by 3.1%, while the model parameters are reduced by 94.33%, the computational cost is reduced by 74.07%, and the model size is reduced by 89.61%. This approach successfully achieves the goal of model lightweighting without compromising accuracy, providing an efficient and lightweight solution for foreign object detection in edge devices.