基于Faster R-CNN多特征融合的人类活动检测网络  

Human activity detection network based on Faster R-CNN multi-feature fusion

作　　者：刘登生 陈利霞[1] 王学文 梁胜梅 LIU Dengsheng;CHEN Lixia;WANG Xuewen;LIANG Shengmei(Guangxi Colleges and Universities Key Laboratory of Data Analysis and Computation, Guilin University of Electronic Technology,Guilin 541004,China;School of Computer and Information Security,Guilin University of Electronic Technology,Guilin 541004,China;School of Mathematics and Computational Science,Guilin University of Electronic Technology,Guilin 541004,China)

机构地区：[1]桂林电子科技大学广西高校数据分析与计算重点实验室,广西桂林541004 [2]桂林电子科技大学计算机与信息安全学院,广西桂林541004 [3]桂林电子科技大学数学与计算科学学院,广西桂林541004

出　　处：《桂林电子科技大学学报》2018年第5期364-368,共5页Journal of Guilin University of Electronic Technology

基　　金：国家自然科学基金(61362021;61661017);广西自然科学基金(2013GXNSFDA019030);广西高校数据分析与计算重点实验室基金(LDAC201704);广西高校中青年教师基本研究能力推广项目(KY2016YB162);桂林电子科技大学研究生教育创新计划(2017YJCX84)

摘　　要：针对网络自身存在的网络特征提取不充分情况,将Faster R-CNN的共享卷积层进行前层特征融合,使得前层有用却不明显的特征可以在后层中被重新利用,达到增强网络细微特征提取的效果,并采用优化非极大值抑制(NMS)方法筛选候选框。实验结果表明,基于Faster R-CNN多特征融合的人类活动检测网络不仅可更完整地检测出人所在的位置,而且对目标漏检情况也有所改善,对活动中人的检测精度比Faster R-CNN网络提高了4.5%。For reasons that Faster R-CNN network can not extract images features fully,the pre-layer characteristics of the Shared convolutional layer of Faster R-CNN are fused,and then,the characteristics of former layer which is useful but not obvious can be used again in the back layer to enhanced the ability of fine feature extraction.Simultaneously,non-maximum value suppression(NMS)method is optimized to select the candidate boxes.Experiments show that the improved network not only has a more complete location to the person,but also improves the leak detection.Compared with the Faster R-CNN,the detection accuracy of our improved network increases 4.5%,which provided a concise solution for an accurate and rapid detection.

关 键 词：人类活动检测 深度学习 特征融合 Soft-NMS 

分 类 号：TN911.72[电子电信—通信与信息系统]