多尺度特征融合的露天矿区道路负障碍检测

Road negative obstacle detection in open-pit mines based on multi scale feature fusion

露天矿区非结构化道路存在的坑洼、道路塌陷等负向障碍,易导致车辆侧翻或陷车,近年来矿用卡车自动驾驶的兴起,使得负向障碍检测变得至关重要。对露天矿区道路负障碍特征进行深入分析,构建了基于机器视觉的轻量化目标检测模型。首先通过现场采样及标注建立露天矿区负向障碍数据集并将其输入到目标检测模型;其次对图像进行归一化处理并使用MobileNetv3网络对图像进行压缩激活,在获得输出特征后进行连续上采样和特征金字塔堆叠,完成多尺度特征提取;最后对多尺度特征进行分类和边界框回归,达到负向障碍检测的目的。模型在特征金字塔模块中引入深度可分离卷积方法,降低网络特征提取和融合的计算量;通过对损失函数和学习率动态优化调整,提高负障碍目标检测精度;在负障碍检测后处理阶段,提出非极大抑制优化算法,改善负障碍被遮挡和检测框定位精度不高的问题。实验结果表明,研究方法能有效识别复杂背景下露天矿区非结构化道路的负障碍,对矿区道路负障碍检测精度、召回率、mAP达到98.86%,89.58%和92.59%,实时检测速度达到47.3 fps。对比主流的目标检测网络,如:Yolov3,RetianNet,SSD,Faster RCNN等也有着更好的综合性能。通过对非极大抑制优化算法的量化分析,相较于传统算法,改进算法可有效改善检测框的定位精度,同时具有良好的适应性。

The existence of negative obstacles such as potholes and road cave-ins on unstructured roads in open pit mines can easily lead to vehicle rollover or trapping,and the development of autonomous mining truck driving in recent years has made negative obstacle detection crucial.In this study,the characteristics of road negative obstacles in open-pit mining area are comprehensively analyzed,and a lightweight object detection model based on machine vision is constructed.Firstly,the negative obstacle data set of open-pit mining area is constructed by field sampling and input into the target detection model.Secondly,the image is normalized and activated by mobileNetv3 network.After obtaining the output features,the continuous up-sampling and feature pyramid stacking are performed to complete multi-scale feature extraction.Finally,the multi-scale features are classified and the bounding box regression is performed to achieve negative obstacle detection.The model introduces a depth-separable convolution method in the feature pyramid module to reduce the computational effort of network feature extraction and fusion.By adjusting the loss function and learning rate dynamically and optimally,the accuracy of negative obstacle target detection is improved.In the post-processing stage of negative obstacle detection,a non-maximum suppression optimization algorithm is proposed in the post-processing stage of negative obstacle detection to solve the problem that the negative obstacles are obscured and the detection frame is not localized with high accuracy.The experimental results show that the research method can effectively identify negative obstacles on unstructured roads in open-pit mining areas in complex backgrounds.The mine road negative barrier detection precision,recall and mAP reach98.86%,89.58%and 92.59%with a real time detection speed of 47.3 fps.It also has a better overall performance than that of mainstream target detection networks such as Yolov3,RetianNet,SSD,Faster RCNN etc.A quantitative analysis of the non-maximal suppression optimization algorithm has also shown that the improved algorithm is effective in improving the positioning accuracy of the detection boxes compared to the traditional algorithm and is also very adaptable.