详细信息
文献类型:期刊文献
中文题名:基于相邻特征融合的目标检测
英文题名:Object Detection Based on Neighbour Feature Fusion
作者:李亚泽[1,2];刘宏哲[1,2]
第一作者:李亚泽
机构:[1]北京联合大学北京市信息服务工程重点实验室,北京100101;[2]北京联合大学机器人学院,北京100101
第一机构:北京联合大学北京市信息服务工程重点实验室
年份:2021
卷号:48
期号:12
起止页码:264-268
中文期刊名:计算机科学
外文期刊名:Computer Science
收录:CSTPCD;;北大核心:【北大核心2020】;CSCD:【CSCD_E2021_2022】;
基金:国家自然科学基金(61871039,61906017,61802019);北京市教委项目(KM202111417001,KM201911417001);视觉智能协同创新中心项目(CYXC2011);北京联合大学学术项目(ZK80202001,202011417004,202011417005)。
语种:中文
中文关键词:深度学习;目标检测;计算机视觉;特征融合;智能驾驶
外文关键词:Deep learning;Object detection;Computer vision;Feature fusion;Autonomous driving
摘要:随着智能驾驶领域的发展,人们对目标检测的精度要求越来越高,尤其是针对高速行驶时对距离较远的小目标的检测和低速行驶时对密集目标的检测。在当前的两阶段检测框架的特征融合部分,使用bottom-up的双向融合方法虽然能够更有效地对大目标进行语义信息和位置信息的特征融合,但会给几个或几十个像素的小目标造成很大的信息损失。当检测网络特征融合部分使用top-down的单向融合方法时,则对大目标检测的效果欠佳。为此,文中提出了相邻特征融合(Neighbour Feature Pyramid Network,NFPN)方法、Double RoI(Region of Interest)方法和递归特征金字塔(Recursive Feature Pyramid,RFP)的方法。以Faster RCNN 50为基准,同时使用提出的NFPN,Double RoI和RFP后,在Lisa交通数据集中平均精度(mAP)提升了2.6个百分点。在VOC2007数据集上,以VOC07+12 train数据集为训练集,VOC2007 test为测试集,以Faster RCNN101为基准,同时使用提出的3个模型,mAP提升了6个百分点,同时小、中、大目标的精度也得到提高。
With the development of intelligent driving,the precision requirements for target detection are getting higher and higher,especially for small targets that are far away.In the neck of two-stage object detection network,although the feature fusion of semantic information and location information is more effective for large targets if the bottom-up fusion method is used,it will cause big information loss to small targets.To address this problem,we propose neighbor feature pyramid networks(NFPN)method of feature fusion of neighbor layers,the Double RoI(Region of Interest)method to fuse the FPN and NFPN features,and the recursive feature pyramicl(RFP)method.Using Faster RCNN 50 as the benchmark,the mean average precision(mAP)of our model in the Lisa data set has increased by 2.6%while using NFPN,Double RoI and RFP.On the VOC2007 data set,using the VOC07+12 train data set for training,VOC2007 test as the test set,and Faster RCNN101 as the baseline,the mAP of our model both used NFPN,Double RoIE and RFP has increased by 6%,and the object detect accuracy of large,medium and small targets is improved at the same time.
参考文献:
正在载入数据...