基于多特征融合的露天矿区道路负障碍检测

Negative obstacle detection on open pit roads based on multi-feature fusion

随着智慧矿山概念的逐步落实,智能化、无人化逐渐在矿区落实,露天矿卡车无人驾驶日益成为矿山智能化建设的主要内容,为解决露天矿区复杂多变的道路因坑洼、塌陷等路面小部分下陷出现的非规则负障碍而导致矿区无人车、重载卡车侧翻等安全难题,提升矿区安全驾驶系数,提出一种多特征融合的露天矿区道路负障碍检测方法。该方法使用BiFPN特征融合模块,提高小型负障碍检测权重占比;引入空间和通道双注意力机制提高对负障碍边缘的特征提取和特征融合能力,从而提高对道路小尺度负障碍的检测精度;采用SIoU Loss作为模型边界框损失函数并使用K-means++方法优化Anchor以提高负障碍检测模型的收敛速度和边界框定位效果,并基于遗传算法优化超参数让模型更贴合矿区场景,最终实现对矿区道路负障碍的快速精准识别。实验表明该检测模型能快速准确识别复杂背景下露天矿区道路负障碍目标,对道路负障碍目标的检测精度、召回率、平均精确度均值分别达到了96.9%、89.9%、95.3%,且该模型大小仅有12.7 MB。对比其他主流检测网络,该网络模型更适合复杂环境下露天矿区非结构化道路行驶安全需求,且该检测模型的鲁棒性好,可适配于多种情况的露天矿区,为实际环境复杂多变的露天矿区非结构化道路负向障碍检测提供了可行的方法,为露天矿无人卡车安全生产运输提供安全预警。

With the gradual implementation of intelligent mine concept,intelligence and unmanned operation are gradually implemented in mining area.Unmanned driving of open pit mine trucks is increasingly becoming the main focus of intelligent mine construction.In order to solve the safety problems of the overturn of unmanned vehicles and heavy-duty trucks due to irregular negative obstacles appearing in some parts of road surface such as potholes and collapses in open pit mines,and to improve the safe driving coefficient in mines,a multi-feature fusion method of detecting negative obstacles in open pit mine roads is proposed.The method uses the BiFPN feature fusion module to improve the weight proportion of small-scale negative obstacle detection,introduces the spatial and channel dual attention mechanism to improve the feature extraction and feature fusion ability of negative obstacle edges,so as to improve the detection accuracy of small-scale negative obstacles on the road.Also,the SIoU Loss is adopted as the loss function of the model bounding box,the Anchor by using the K-means++method is used to improve the convergence speed and boundary frame localiza-tion effect of the obstacle detection model,the hyperparameters are optimized based on genetic algorithm to make the model more suitable for the mine scene,and finally the fast and accurate recognition of negative obstacles on the mine road is realized.The experiments show that the detection model can quickly and accurately identify the negative road obstacle targets in the complex background of the open pit mine,and the detection accuracy,recall rate,and mAP of the negative road obstacle targets reach 96.9%,89.9%,and 95.3%,respectively,and the size of the model is only 12.7 MB.Compared with other mainstream detection networks,the network model is more suitable for the safety needs of unstruc-tured road driving in open pit mining areas under complex environment,and the robustness of the detection model is good,which can be adapted to a variety of situations in open pit mining areas,providing a feasible method for the detection of negative obstacles on unstructured roads in open pit mining areas where the actual environment is complex and variable,and providing some safety warnings for the safety of unmanned trucks in open pit mines.