井下无人驾驶电机车行驶场景中多目标检测研究

Research on multi-object detection in driving scene of underground unmanned electric locomotive

  • 摘要: 目前煤矿井下无人驾驶有轨电机车在行驶过程中,对轨道中的石块及其他小型障碍物的识别存在检测速度慢、检测精度低,且对于重叠目标,易造成漏检、错检等问题。针对上述问题,提出了一种井下电机车多目标检测模型−SE−HDC−Mask R−CNN模型。该模型基于Mask R−CNN进行改进,通过在主干特征提取网络ResNet的残差块中嵌入压缩−激励(SE)模块,学习各个通道的重要程度和相互联系,增强网络对特征的选择和捕获能力;将残差块中卷积核大小为3×3的标准卷积替换成混合空洞卷积(HDC),在不改变特征图大小、不增加参数计算量的前提下,通过增加卷积核处理数据时各值之间的距离达到增大感受野的目的。实验结果表明:SE−HDC−Mask R−CNN模型可有效提取轨道、电机车、信号灯、行人和石块目标,在井下电机车多场景运行数据集上的平均准确率均值为95.4%,平均掩码分割精度为88.1%,平均边界框交并比为91.7%,相较于Mask R−CNN模型均提升了0.5%,对信号灯、石块(小目标)的检测精度分别提升了0.7%和4.1%;SE−HDC−Mask R−CNN模型的综合性能优于YOLOV2,YOLOV3−Tiny,SSD,Faster R−CNN等模型,可有效解决小目标漏检问题;SE−HDC−Mask R−CNN模型在煤巷直轨、弯轨、黑暗环境、多目标重叠等场景下均可有效实现目标检测,具有一定泛化能力及较高鲁棒性,基本满足无人驾驶电机车障碍物检测需求。

     

    Abstract: At present, there are some problems in the identification of stones and other small obstacles in the track during the driving of unmanned underground electric locomotive in coal mines, such as slow detection speed, low detection precision, and easy to cause missing detection and wrong detection for overlapping objects. In order to solve the above problems, a multi-target detection model (SE-HDC-Mask R-CNN) for underground electric locomotive is proposed. The model is improved on the basis of Mask R-CNN. By embedding a squeeze-and-excitation (SE) module in the residual block of the backbone feature extraction network ResNet, the importance and interrelation of each channel are learned. The capability of feature selection and capture of the network is enhanced. The standard convolution with a kernel size of 3×3 in the residual block is replaced with hybrid dilated convolution (HDC). On the premise of not changing the size of the feature image and not increasing the amount of parameter calculation, the receptive field can be increased by increasing the distance between the values when the convolution kernel processes the data. The experimental result show that the SE-HDC-Mask R-CNN model can effectively extract track, electric locomotive, signal light, pedestrian and stone objects. The average precision rate on the multi scene operation data set of underground electric locomotive is 95.4%, the average mask segmentation precision is 88.1%, the average bound box intersection ratio is 91.7%, the three indicators are all improved by 0.5% compared with the Mask R-CNN model. The detection precision of signal light and stone (small objects) is improved by 0.7% and 4.1% respectively. The comprehensive performance of SE-HDC-Mask R-CNN model is better than that of YOLOV2, YOLOV3-Tiny, SSD and Faster R-CNN model. The SE-HDC-Mask R-CNN model can effectively solve the problem of missing detection of small objects. The SE-HDC-Mask R-CNN model can effectively realize object detection in coal roadway straight track, curved track, dark environment, multi-object overlapping and other scenarios. It has certain generalization capability and high robustness, and basically meets the requirements of unmanned electric locomotive obstacle detection.

     

/

返回文章
返回