光学精密工程2025，Vol.33Issue(14)：2278-2290,13.DOI:10.37188/OPE.20253314.2278
基于头部与整体信息联合的行人检测算法
Pedestrian detection algorithm based on joint head and overall information
摘要
Abstract
In crowded scenes,factors such as occlusion of the pedestrian body and varying pedestrian scales lead to a decrease in the precision of the detector.Since,the pedestrian head tends to be lightly oc-cluded and it can be used to assist detection.In this regard,a pedestrian-detection algorithm based on joint head and overall information was proposed.First,a feature-extraction network built upon dense connec-tions and enhanced fusion was designed to strengthen multi-scale feature extraction and to improve the net-work's sensitivity to pedestrians of various scales.Second,the sampling mechanism of the region-propos-al-network module was optimized by introducing a non-uniform hard-example-mining strategy that discrimi-nated according to the occlusion-overlap rate;this strategy concentrated on heavily occluded hard samples and enhanced the network's adaptability to occlusion.Next,a joint head-and-body detection strategy was constructed,and the post-processing stage was refined so that head-detection results could recover body boxes that had been erroneously suppressed by occlusion,thereby reducing the missed-detection rate.Meanwhile,the loss function was further optimized by incorporating the characteristics of the joint detec-tion box,so that mis-detections and missed detections caused by occlusion were alleviated.Finally,the ef-fectiveness of the proposed algorithm was verified through experiments.The results show that the pro-posed algorithm reduces the log-average miss detection rate by 5.7％and improves the average precision by 4％on the CrowdHuman dataset with a higher degree of occlusion,and reduces the log-average miss detection rate by 2.4％and 2.1％on the two small-scale subsets of the TJU-DHD-pedestrian dataset,which effectively enhances the detection capability of both occluded pedestrian targets and multi-scale pe-destrian targets.关键词
行人检测/联合检测/多尺度特征融合/难例挖掘/后处理优化Key words
pedestrian detection/joint detection/multi-scale feature fusion/hard example mining/post-processing optimization分类
信息技术与安全科学引用本文复制引用
马晞茗,李宁,吴迪,刘一豆,于祥跃,李峥..基于头部与整体信息联合的行人检测算法[J].光学精密工程,2025,33(14):2278-2290,13.基金项目
国家自然科学基金(No.61705219) （No.61705219）
中国科学院青年创新促进会会员(No.2022216) （No.2022216）
