Abstract: In the process of coal folw transportation on mine conveyors, there are different sizes and shapes of rock bolt and large coal gangue, so it is difficult to extract image feature imformation, and the inspection effect of traditional objecet detection algorithm is not deal. To solve this problem, a foreign object detection algorithm based on TDConv and a unified attention detection head is proposed.This algorithm designs a TDConv convolution module by combining parallel convolution methods, effectively preserving the original information of image features and assisting deeper convolutional layers in extracting detailed information.An unified attention module is incorporated into the detection head to effectively extract and recognize feature information from objects of different sizes and spatial positions. A dataset of 100 000 images (MFID) of mining foreign objects was created based on various underground coal mine scenarios, providing resources for in-depth research and practical application of foreign object detection during coal flow transportation. Experimental results demonstrate that this algorithm improves the mean Average Precision (mAP) by 2.1% compared to the YOLOv5 object detection algorithm on the MFID mining dataset.Furthermore, it effectively reduces the parameter count of the foreign object detection network model while maintaining high detection accuracy, resulting in a more lightweight network structure suitable for magrinal computing devices in underground coal mines.