重庆科技大学学报（自然科学版）2025，Vol.27Issue(6)：69-79,11.DOI:10.19406/j.issn.2097-4531.2025.06.008
基于自适应锚框分配的行人检测方法
Pedestrian Detection Method Based on Adaptive Anchor Box Assignment
摘要
Abstract
To address the issue of low accuracy in pedestrian detection caused by significant variations in target scales and the imbalance between positive and negative samples in single-stage multi-layer target detection algorithms,a pedestrian detection method based on adaptive anchor box assignment is proposed.Firstly,replace the backbone network of the SSD algorithm with ResNet-50,leveraging the deep residual structure to address the vanishing gradi-ent problem,and introduce additional convolutional layers to fuse features through 1×1 convolution and bilinear interpolation,optimizing information flow with the skip connections of residual modules,thereby enhancing the de-tection capability for multi-scale objects,especially small ones.Secondly,a lightweight channel attention feature fusion pyramid network-LE-FPN is constructed which promotes cross-layer feature interaction and enhances the net-work's ability to focus on key features through lightweight FPN design and ECA module embedding.Finally,an adaptive anchor box assignment strategy is designed that dynamically selects positive samples based on the statistical characteristics of the targets.This approach avoids the limitations of fixed IoU thresholds,thereby improving multi-scale object detection performance,and enhancing the model's generalization ability.Experimental results on the JAAD dataset indicate that the overall prediction accuracy of this method is 12.04 percentage points higher than that of the benchmark algorithm.关键词
自适应锚框分配/轻量化特征金字塔网络/轻量级注意力机制/行人检测Key words
adaptive anchor box assignment/lightweight feature pyramid network/lightweight attention mecha-nism/pedestrian detection分类
信息技术与安全科学引用本文复制引用
雷亮,赵锦,刘学涵,陈毅,徐山雯,周华勇,陈小庆..基于自适应锚框分配的行人检测方法[J].重庆科技大学学报（自然科学版）,2025,27(6):69-79,11.基金项目
重庆市教委科学技术研究项目"基于YOLOPose姿态特征检测融合时序的独居老人跌倒识别算法研究"(KJQN202303305) （KJQN202303305）
2021年重庆市属本科高校与中科院所属院所合作项目"工业互联网内生安全关键技术研究与协同创新"(HZ2021015) （HZ2021015）
