井下视频行人检测方法  

Underground video pedestrian detection method

作　　者：李现国[1,2] 李斌 刘宗鹏 冯欣欣 刘晓 宋金水 张磊 LI Xianguo;LI Bin;LIU Zongpeng;FENG Xinxin;LIU Xiao;SONG Jinshui;ZHANG Lei(School of Electronic and Information Engineering,Tianjin Polytechnic University,Tianjin 300387,China;Tianjin Key Laboratory of Optoelectronic Detection Technology and System,Tianjin 300387,China;Shandong New Dragon Energy Co.,Ltd.,Heze 274918,China)

机构地区：[1]天津工业大学电子与信息工程学院,天津300387 [2]天津市光电检测技术与系统重点实验室,天津300387 [3]山东新巨龙能源有限责任公司,山东菏泽274918

出　　处：《工矿自动化》2020年第2期54-58,共5页Journal Of Mine Automation

基　　金：天津市重点研发计划科技支撑重点项目(18YFZCGX00930)

摘　　要：针对现有基于深度学习的行人检测方法存在计算量较大、检测效率严重依赖硬件性能等问题,对基于SSD网络的行人检测方法进行改进,设计了一种基于DenseNet网络的轻量级卷积神经网络作为SSD网络的基础网络,以满足井下视频行人实时检测需求,并设计了基于ResNet网络的辅助网络,以增强特征表征能力,提高行人检测准确性。将基于改进SSD网络的井下视频行人检测方法部署在嵌入式平台Jetson TX2上进行实验,结果表明该方法对井下视频中行人的检测准确率为87.9%,针对井下行人低密度场景的检测准确率近100%,且运算速度达48帧/s,约为基于SSD网络的行人检测方法的4.4倍,满足井下行人实时检测需求。For problems of existing pedestrian detection methods based on deep learning such as large computation,detection efficiency relying on hardware performance heavily and so on,the pedestrian detection method based on SSD network was improved.A lightweight convolutional neural network based on DenseNet network is designed as basic network of SSD network to meet real-time detection requirements of underground video pedestrian,and an auxiliary network based on ResNet network is designed to strengthen feature extraction ability and improve correctness of pedestrian detection.An underground video pedestrian detection method based on the improved SSD network has been arranged in embedded platform Jetson TX2 for experiments.The results show that detection accuracy rate of the method for underground video pedestrian is 87.9%as well as nearly 100%for underground low-density pedestrian scene,and calculation speed achieves 48 frames per second,which is about 4.4 times as quick as the pedestrian detection method based on SSD network and meets real-time detection requirements of underground pedestrian.

关 键 词：井下行人检测 视频监控 深度学习 SSD网络 卷积神经网络 

分 类 号：TD67[矿业工程—矿山机电]