基于OC&PGMF的弱监督行人检测方法  

OC&PGMF for weakly supervised pedestrian detection

作　　者：曹鎏 徐巧玉[1] CAO Liu;XU Qiao-yu(School of Mechatronics Engineering,Henan University of Science and Technology,Luoyang 471000,China)

机构地区：[1]河南科技大学机电工程学院,河南洛阳471000

出　　处：《计算机工程与设计》2024年第9期2725-2732,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(51205108)。

摘　　要：为解决弱监督行人检测训练过程中数据收敛到局部最优解和缺少回归能力问题,提出一种基于改进的在线学习与伪真值挖掘过滤算法的弱监督行人检测方法。采用WSDDN作为基础的弱监督检测器,通过优化OICR在线学习精细化策略,增加行人检测召回率并覆盖行人完整的位置,改善算法收敛到局部最优解的问题;基于邻域融合原理,通过设计伪真值挖掘过滤算法优化伪真值,进行全监督行人检测器训练,提高弱监督行人检测器的回归能力。实验结果表明,所提弱监督检测方法在VOC2007上实现了21.3%的mAP准确率,高于经典的弱监督行人检测方法(PCL)3.5%mAP准确率,验证了其有效性。To address the challenges of data converging to a local optimal solution and regression ability lacking in weakly supe-rvised pedestrian detection during training stage,a pedestrian detection method was proposed based on the improved online lear-ning(OC)and pseudo ground truth mining filtering(PGMF)algorithm.An improved online learning was plugged into base weakly supervised detector(WSDDN),which increased the recall rate of pedestrian detection and completely covered pedestrian regions.PGMF algorithm was designed to optimize initial pseudo ground truth and fully supervised pedestrian detectors were trained using updated pseudo ground truth,which improved regression ability of weakly supervised pedestrian detectors.Extensive experiments demonstrate the effectiveness of the proposed method in weakly supervised pedestrian detection,and achieves the state-of-the-art 21.3%in mAP on PASCAL VOC2007 benchmark,surpassing PCL by 3.5%absolutely.

关 键 词：行人检测 弱监督学习 在线学习 伪真值 挖掘过滤 局部最优解 回归能力 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]