基于加权双向特征融合的矿区道路落石检测  

Road Rockfall Detection in Mining Area Based on Weighted Bidirectional Feature Fusion

作　　者：顾清华[1,2] 杜艺凡 李萍丰 王丹 GU Qinghua;DU Yifan;LI Pingfeng;WANG Dan(School of Resources Engineering,Xi’an University of Architecture and Technology,Xi’an 710055,Shaanxi,China;Xi’an Key Laboratory of Intelligent Industry Perception Computing and Decision Making,Xi’an 710055,Shaanxi,China;Hongda Blasting Engineering Group Ltd.,Co.,Guangzhou 511300,Guangdong,China)

机构地区：[1]西安建筑科技大学资源工程学院,陕西西安710055 [2]西安市智慧工业感知计算与决策重点实验室,陕西西安710055 [3]宏大爆破工程集团有限责任公司,广东广州511300

出　　处：《黄金科学技术》2023年第6期953-963,共11页Gold Science and Technology

基　　金：国家自然科学基金项目“金属露天矿无人驾驶多工序多目标协同智能调度方法研究”(编号:52074205);陕西省杰出青年基金项目“时空路况下金属露天矿无人驾驶多车协同智能调度集成建模”(编号:2020JC-44)联合资助。

摘　　要：为了解决当前露天矿区非结构化路面落石检测存在的环境复杂、落石尺寸差异较大以及落石与非结构化路面颜色相近而造成难以精准识别的问题,提出了一种基于加权双向特征融合的矿区道路落石检测模型。首先,通过加入SimAM注意力机制有效抑制背景环境的干扰;其次,使用加权双向特征金字塔(BiFPN)结构实现多尺度特征融合,增强模型对于不同尺寸落石的特征提取能力;最后,引入轻量级卷积GSConv模块,通过减少模型计算量来提升模型检测速度。试验结果表明,该算法的检测精度均值达到92.8%,检测速度达到63.1FPS,能够实现矿区非结构化路面落石的实时高精度检测,为无人矿卡的安全行驶提供了保障。With the booming development of big data and Internet of Things technology,traditional mines have developed to smart mines and intelligent mines,and unmanned technology has been gradually applied to mining areas.In order to solve the problem that the rockfall detection of unstructured road in open-pit mine area has complex environment,large difference in rockfall size and similar color between rockfall and unstructured road surface,a rockfall detection model of mining road based on weighted bidirectional feature fusion was proposed.First,the SimAM attention mechanism is added to the backbone network,this attention mechanism is different from the previous channel attention mechanism and spatial attention mechanism,it can effectively eliminate the interference of the background environment without adding additional parameters,so that the model can focus more on the target characteristics of rockfall.Second,the weighted bidirectional feature pyramid(BiFPN)structure was used to realize multi-scale feature fusion in the neck.Since the PANet structure in the YOLOv5s network model only adds or splice the characteristics of the pyramid structure in the melting process,the bidirectional feature weighting was combined with the bidirectional feature of the weight and adaptive adjustment to ensure that the network model attaches proper importance to the rock ebaissees different sizes and different levels and realizes the addition between the low-level position information and high-level semantic information for multiple cross-layer weighted feature fusion,thus enhancing the feature extracion ability of the model for rockfall of different sizes.Finally,the lightweight convolution GSConv module was introduced into the col,which can be used to process function cards at this time,not only reducing redundant information,but also avoiding compression.The GSConv lightweight convolution module is based on deep separable convolution(DSC),ordinary convolution(SC)and channel shuffle operation,which improves the detection speed of the m

关 键 词：矿区道路 无人驾驶 机器视觉 落石检测 多尺度特征融合 YOLOv5模型 

分 类 号：TD57[矿业工程—矿山机电] TP391.41[自动化与计算机技术—计算机应用技术]