现代电子技术2025，Vol.48Issue(11)：34-41,8.DOI:10.16652/j.issn.1004-373x.2025.11.006
基于Crowd-RetinaNet的拥挤行人检测
Crowd-RetinaNet based detection of the crowded
摘要
Abstract
An efficient network model for congested pedestrian detection is designed on the basis of the RetinaNet.The coordinate attention feature fusion(CAFF)is used to carry out cross-layer feature fusion to realize high-quality semantic and positional detail information interaction between scale features,so as to improve the performance of feature fusion.The task-aware head(TaHead)is introduced to improve the representation ability of object detection head,so as to improve the performance of object detection.In combination with the multi-instance prediction(MIP)algorithm implemented by the CrowdDet,the earth mover's distance loss(EMDLoss)algorithm is used to train the model,and the Set NMS,as a post-processing method,is used to suppress multiple redundancy detection results effectively,so as to overcome the non-maximum suppression(NMS)algorithm missing detection of occluded objects,and finally a Crowd-RetinaNet crowded pedestrian detection model is designed.The Crowd-RetinaNet was trained on the CrowdHuman data set and its performance test was performed on the CrowdHuman validation set.In comparison with the basic model RetinaNet,the AP(average precision)and MR-2 of the Crowd-RetinaNet is improved by 1.80%and 3.32%,respectively.In addition,the experiment of pedestrian detection for high performance is completed in the crowded scene of campus.To sum up,it is of great significance to improve the overall performance of smart home.关键词
行人检测/坐标注意力特征融合/信息交互/多实例预测/注意力机制/抑制算法Key words
pedestrian detection/coordinate attention feature fusion/information exchange/MIP/attention mechanism/suppression algorithm分类
信息技术与安全科学引用本文复制引用
韩鼎,喻春雨,童亦新,张俊..基于Crowd-RetinaNet的拥挤行人检测[J].现代电子技术,2025,48(11):34-41,8.基金项目
南京邮电大学校企合作项目(2018外002,2019外157) （2018外002,2019外157）
