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.Exten-sive 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.
维普
万方数据