Abstract: Pedestrian detection is an active topic of research in computer vision. In order to implement the detection of pedestrian in complicated environments, Region-based Fully Convolutional Networks（R-FCN） is proposed to pedestrian detection in this study. Addressed to occlusion, background interference and small objects in pedestrian detection, this paper changes the searching mechanism of R-FCN, and introduces area division for target pedestrian（upper part, lover part of pedestrian） and reinforcement learning strategy of background interference, to strengthen the learning of occlusion and samples similar to the background. On this basis, the second-time classification on the output of the R-FCN is operated. The experimental results indicate that the improvement of R-FCN effectively solves the problem of missing and misclassification from traditional R-FCN classifier under the condition of occlusion, background interference and small objects.

Key words: Region-based Fully Convolutional Networks（R-FCN）, occlusion, background interference, second-time classification

摘要： 行人检测是计算机视觉中的研究热点，为了实现复杂场景下的行人检测，将基于区域的全卷积网络（Region-based Fully Convolutional Networks，R-FCN）引入行人检测中。针对行人检测中的遮挡、背景混淆干扰、小目标这三个难点，修改了R-FCN的搜索机制，引入目标行人的区域划分（上下半身）和背景混淆干扰行人的强化学习策略，加强了对遮挡行人和背景相似行人的学习。并在此基础上，对R-FCN的输出进行二次分类学习。实验结果表明，通过对R-FCN的改进，可有效地缓解行人遮挡、背景混淆干扰和小目标条件下，传统R-FCN网络的漏报和误判问题。

关键词: 基于区域的全卷积网络（R-FCN）, 遮挡, 背景混淆干扰, 二次分类