Technologies for unmanned driving hold a prominent place in the open-pit mining industry. Intelligent detection technology is crucial for applications involving unmanned driving. Nevertheless, object detection accuracy is somewhat low in the complex mine scenario. This work generated an open-pit mine dataset and examined it from various angles in order to achieve good detection accuracy in the open-pit mine scenario. Second, in order to concentrate on the area of interest, we suggested a hybrid attention module made up of single-head and multi-head attention modules. In addition, an additional branch was created using the CNN network in conjunction with the CBAM block to improve the knowledge about small objects. In order to reduce computation parameters and achieve real-time detection, depthwise separable convolution was used in place of the convolutional layer. Lastly, to improve detection generalization, L2 regularization was used to control the loss function. Our suggested network could achieve the 89.14% mAP. Additionally, the detection result shows that the mAP of our suggested network is 2.1% greater than that of the traditional YOLOv4 model. For better YOLOv4, the average F1 score was also tested at 0.87. The heatmap also suggests that our suggested network may be able to enhance target expression and capture rich discriminative feature representations. Results from experiments confirm that our suggested method could function well in a challenging open-pit mine situation.
Wang et al. (Fri,) studied this question.