Abstract: In order to solve the problems of harsh underground environment, poor lighting, mixed background and multi-scale pedestrian, this paper proposes a pedestrian detection method of underground coal mine based on improved Faster RCNN. Deep convolutional neural network can replace traditional manual design feature to extract features automatically from images. Based on the Faster RCNN algorithm, RPN（Region Proposals Network） structure is improved and a “pyramid RPN” structure is proposed to solve multi-scale detection problem of pedestrian underground. At the same time, by adding feature fusion technology, the feature maps of different convolution layers are merged to improve the detetion performance for under-mine blur, occlusion and tiny pedestrian. The experimental results indicate that the improved Faster RCNN can effectively solve the pedestrian detection problem of underground coal mine, which obtains 90% detection accurary on the under-mine pedestrian dataset. The improved Faster RCNN algorithm is validated in the VOC 07 benchmark.

Key words: deep learning, Faster RCNN, pedestrian detection

摘要： 针对煤矿井下环境恶劣、光照差、背景混杂、行人模糊、行人多尺度等问题，提出了一种改进的Faster RCNN煤矿井下行人检测方法，使用深度卷积神经网络代替传统的手工设计特征方式自动地从图片中提取特征。利用深度学习通用目标检测框架Faster RCNN，以Faster RCNN算法为基础，对候选区域网络（Region Proposals Network，RPN）结构进行了改进，提出了一种“金字塔RPN”结构，来解决井下行人存在的多尺度问题；同时算法中加入了特征融合技术，将不同卷积层输出的特征图进行融合，增强煤矿井下模糊、遮挡和小目标行人的检测性能。实验结果表明：改进的Faster RCNN可以有效解决井下行人检测问题，在井下行人数据集上获得了90%的检测准确率，并在公测数据集VOC 07上对改进算法进行了验证。

关键词: 深度学习, Faster RCNN, 行人检测