Abstract: In order to solve the problem of slow detection speed of existing deep learning based belt conveyor foreign object detection methods, an improved YOLOv3 model is proposed and applied to coal mine belt conveyor foreign object detection. The model uses the lightweight network DarkNet22-DS as the backbone feature extraction network. DarkNet22-DS replaces the standard convolution with depthwise separable convolution, which reduces the network parameters significantly and improves the feature utilization efficiency by composite residual blocks. By introducing weighted bi-directional feature pyramid networks and dual-scale output, the model improves the feature fusion network and enhances the model's detection efficiency of large foreign objects. The complete intersection ratio loss function is used as the target box regression loss function, and the correlation between the target box information is fully utilized to improve the convergence speed and detection accuracy of the model. The improved YOLOv3 model is deployed on the embedded platform Jetson Xavier NX for coal mine belt conveyor foreign object detection experiments. The results show that compared with the YOLOv3 model, the weight file size of the improved YOLOv3 model is reduced by 91.4%, and the amount of model parameters is reduced significantly. The detection speed is increased by 16 times, reaching 30.7 frames/s. The performance meets the real-time detection requirements of foreign objects in coal mine belt conveyors.