改进YOLOv3网络结构的遮挡行人检测算法  

Occluded Pedestrian Detection Algorithm Based on Improved Network Structure of YOLOv3

作　　者：刘丽[1,2] 郑洋[1,2,3] 付冬梅 LIU Li;ZHENG Yang;FU Dongmei(School of Automation and Electrical Engineering,University of Science and Technology Beijing,Beijing 100083;Beijing Engineering Research Center of Industrial Spectrum Imaging,University of Science and Technology Beijing,Beijing 100083;Shunde Graduate School,University of Science and Technology Beijing,Foshan 528399)

机构地区：[1]北京科技大学自动化学院,北京100083 [2]北京科技大学北京市工业波谱成像工程技术研究中心,北京100083 [3]北京科技大学顺德研究生院,佛山528399

出　　处：《模式识别与人工智能》2020年第6期568-574,共7页Pattern Recognition and Artificial Intelligence

基　　金：北京科技大学中央高校基本科研业务费专项资金(No.FRF-BD-19-002A)资助。

摘　　要：针对YOLOv3算法在监控视频行人检测中对遮挡目标漏检率较高的问题,文中提出改进YOLOv3网络结构的遮挡行人检测算法.首先在网络全连接层引入空间金字塔池化网络,增强网络的多尺度特征融合能力.然后采用网络剪枝的方式,精简网络冗余结构,避免网络层数加深导致的退化和过拟合问题,同时减少参数量.在走廊行人数据集上进行多尺度训练,获得最优的权重模型.实验表明,文中方法在平均准确率和检测速度上都有所提升.Aiming at high missed detection rates of YOLOv3 for occluded pedestrian in surveillance video,a detection method for occluded pedestrian based on improved network structure of YOLOv3 is proposed.Firstly,the spatial pyramid pooling network is introduced into the fully connected layer to enhance the multi-scale feature fusion capability of the network.Secondly,the network structure pruning is employed to eliminate the network structure redundancy to avoid network degeneration and overfitting problem caused by the deepening of network layers and reduce the amount of parameters.Multi-scale training is performed on the corridor pedestrian dataset to obtain the best weight model.Experimental results indicate the improvement of average accuracy and detection speed of the proposed algorithm.

关 键 词：行人检测 深度学习 YOLOv3 空间金字塔池化网络 网络剪枝 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]