摘要
随着金属露天矿开采深度不断加大,道路运输条件愈发复杂,无人矿车行驶在道路上面临着各种障碍物的安全隐患,因此对无人矿卡障碍物智能检测提出了更高要求。提出了一种融合Swin Transformer与CNN的露天矿车前障碍物智能检测方法,障碍物检测模型需要建立长期依赖关系来处理不断增加的图像数据,Swin Transformer可以关注全局语义信息,有利于长期建模。将Swin Transformer融入YOLOX模型的骨干特征提取网络中,充分利用多头注意力机制,对图像特征进行预处理,在加强特征提取网络中加入CBAM注意力机制模块,使模型在后续的特征提取中能够提取更多的表征信息。该模型使用的数据集均来自实地矿山,并采用数据增强方式进行预处理。经过实地矿山数据对比验证试验,结果表明:该方法能够有效识别背景复杂的金属露天矿区非结构化道路障碍物,检测精度达到91.57%m AP,检测速度达到56.86 fps,具有较好的小目标和多尺度目标检测性能,可以满足无人矿卡在金属露天矿区的高精度检测要求。
With the deepening of metal open-pit mining,the road transportation conditions become more and more complex,and unmanned mining cards driving on the road faces the safety hazards of various obstacles,so the intelligent detection of obstacles for unmanned mine cards has put forward higher requirements.In this paper,a fusion of Swin Transformer and Convolutional Neural Network(CNN)for the intelligent detection of obstacles in front of open-pit mining trucks is proposed.The obstacle detection model needs to establish long-term dependencies to deal with increasing image data,and Swin Transformer can focus on global semantic information,which is beneficial to long-term modeling.The Swin Transformer is incorporated into the backbone feature extraction network of the YOLOX model to make full use of the multi-headed attention mechanism to preprocess image features,and the CBAM attention mechanism module is added to the enhanced feature extraction network to enable the model to extract more representational information in the subsequent feature extraction.The datasets used in the model are all from field mines and are pre-processed using data enhancement. After field mine data comparison and validation experiments,the results show that:the method can effectively identify unstructured road obstacles in metal open-pit mine with a complexbackground,and the detection accuracy reaches 91. 57% mAP,the detection speed reaches 56. 86 fps,with better smalltarget and multi-scale target detection performance,which can meet the unmanned mine card in metal open pit with high accuracydetection requirements.
作者
江松
孔若男
李鹏程
卢才武
章赛
李萌
JIANG Song;KONG Ruonan;LI Pengcheng;LU Caiwu;ZHANG Sai;LI Meng(School of Resource Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China;Sinosteel Maanshan General Institute of Mining Research Co.,Ltd.,Maanshan 243000,China;State Key Laboratory of Safety and Health for Metal Mines,Maanshan 243000,China;Xi′an Key Laboratory for Intelligent Industrial Perception,Calculation and Decision,Xi′an 710055,China;Jidong Cement Tongchuan Co.,Ltd.,Tongchuan 727100,China;School of Management,Xi′an University of Architecture and Technology,Xi′an 710055,China)
出处
《金属矿山》
CAS
北大核心
2023年第5期228-236,共9页
Metal Mine
基金
国家自然科学基金项目(编号:52104146)
陕西省自然科学基金项目(编号:2021JQ-509)
中国博士后科学基金项目(编号:2022M722925)
陕西省社会科学基金项目(编号:2020R005)。
