School of Resource Engineering, Xi'an University of Architecture and Technology
Xi'an Key Laboratory of Industrial Perceptual Computing and Decision, Xi'an University of Architecture and Technology
The Department of Computer, Mathematics and Physical Science Sul Ross State University, Alpine

露天矿环境特殊，道路场景复杂多变，在光照不足时会导致矿区道路多目标识别不清、定位不准，进而影响检测效果，给矿区无人矿用卡车的安全行驶带来严重安全隐患。目前的道路障碍物检测模型不能有效解决矿区暗光环境对模型检测效果的影响，同时对矿区小目标障碍物的识别也有较大误差，不适用于矿区特殊环境下障碍物的检测与识别。针对上述问题，提出了一种基于双曲嵌入的露天矿区暗光环境下多目标检测模型。首先，在模型的图像预处理阶段引入卷积神经网路Retinex−Net对暗图像进行增强，提高图像清晰度；然后，针对数据集中特征过多而无重点偏好的问题，在加强特征提取部分添加全局注意力机制，聚集3个维度上更关键的特征信息；最后，在检测模型预测阶段引入双曲全连接层，以减少特征丢失，并防止过拟合现象。实验结果表明：① 基于双曲嵌入的露天矿区暗光环境下道路多目标检测模型不仅对露天矿区暗光环境下的大尺度目标具有较高的分类与定位精度，对矿用卡车及较远距离的小尺度目标即行人也可准确检测与定位，能够满足无人矿用卡车在矿区特殊环境下驾驶的安全需求。② 模型的检测准确率达98.6%，检测速度为51.52 帧/s，较SSD、YOLOv4、YOLOv5、YOLOx、YOLOv7分别提高20.31%，18.51%，10.53%，8.39%，13.24%，对于矿区道路上的行人、矿用卡车及挖机的检测精度达97%以上。

The environment of open-pit mines is distinctive, and the conditions of the roads in them are complex and constantly changing. Insufficient lighting in the area being mined can make it challenging to identify and position multiple targets on the roads. This, in turn, affects the results of detection and poses serious risks to the safe operation of uncrewed mining trucks.Currently available models to identify obstacles on roads cannot accommodate the impact of poor lighting, and thus, yield inaccurate results. They also have significant shortcomings in identifying small obstacles in the mining area. In this study, we develop a multi-target model of detection for the dark/light environment of an open-pit mine using hyperbolic embedding to address the above-mentioned issues. We introduce the Retinex-Net convolutional neural network to the image preprocessing stage of the detection model to enhance dark images and improve their clarity. To address the issue of an excessively large number of features in the dataset without a clear preference for focus, we incorporate the global attention mechanism into the improved process of feature extraction. This enabled the collection of critical feature-related information in three dimensions. Finally, we incorporate a fully connected hyperbolic layer into the prediction stage of the model to minimize feature loss and prevent overfitting. The results of experiments to verify the proposed model showed that ① it could reliably classify and accurately identify large-scale targets in the low-light environment of the open-pit mining area, and was able to highly accurately identify mining trucks and small vehicles over long distances. It could also accurately identify and locate scaled targets, including pedestrians, such that this satisfies meeting the safety-related requirements of uncrewed mining trucks operating in diverse environments.② The model achieved an accuracy of detection of 98.6% while maintaining a speed of 51.52 frames/s, where this was 20.31%, 18.51%, 10.53%, 8.39%, and 13.24% higher than the accuracies of the SSD, YOLOv4, YOLOv5, YOLOx, and YOLOv7, respectively. Its accuracy of detection of pedestrians, mining trucks, and excavators on mining roads exceeded 97%.

国家自然科学基金项目（52074205）；陕西省自然科学基金杰青项目（2020JC-44）。

顾清华，苏存玲，王倩，等. 基于双曲嵌入的露天矿区暗光环境下道路多目标检测模型[J]. 工矿自动化，2024，50（1）：49-56, 114.

GU Qinghua, SU Cunling, WANG Qian, et al. A multi-target road detection model in a low-light environment in an open-pit mining area based on hyperbolic embedding[J]. Journal of Mine Automation，2024，50（1）：49-56, 114.