Detection of small objects in open-pit coal mine stopes using UAV remote sensing images

Open-pit coal mine stopes witness frequent vehicle accidents due to their complex terrains. Therefore, accurately positioning vehicles in the stopes is critical to safe coal mining. To overcome the inaccurate positioning of small targets in the mining areas using Unmanned aerial vehicle (UAV) remote sensing images, this study proposed an improved YOLOv7 model for object detection. First, to expedite the reasoning of the YOLOv7 network, the ELAN module in the original YOLOv7 model was improved by introducing partial convolution. Based on this, the eSE channel attention mechanism was combined to form the PConv-eSE convolution attention module. The purpose is to enhance the network’s ability to extract the features of small targets and reduce the impact of background information. Finally, the loss function of the Normalized Wasserstein Distance (NWD) metric was used to further optimize the network and improve the accuracy. The improved YOLOv7 model was verified through experiments using a vehicle dataset of a mining area stope. The results show that the P_mav value of the improved YOLOv7 model reached 94.5%, which was 7.2% higher than the original model, thus effectively overcoming the missing of small targets from remote sensing images in the detection using the original network. This study will provide a theoretical basis for the positioning of small objects in an open-pit coal mining area using UAVs.