重庆科技大学学报（自然科学版）2025，Vol.27Issue(3)：57-69,13.DOI:10.19406/j.issn.2097-4531.2025.03.006
基于多尺度特征融合的改进YOLOv8n车辆与行人目标检测算法
Improved YOLOv8n Vehicle and Pedestrian Object Detection Algorithm Based on Multi-Scale Feature Fusion
摘要
Abstract
BF-YOLOv8n,a multi-scale feature fusion-based vehicle and pedestrian target detection algorithm,is proposed to address the issue of detection accuracy degradation in the YOLOv8n algorithm caused by significant var-iations in image sizes and target scales.Firstly,CA-SPPF,an advanced selective feature extraction module,is de-signed in the spatial pooling layer to achieve multi-level fusion and enhance the extraction capabilities of features at different scales.Secondly,a 160×160 pedestrian small-target detection head is designed to improve the model's detection accuracy for small targets.Thirdly,a bidirectional fusion feature pyramid based on four detection heads(BiFPN-4H),is proposed to enhance the model's adaptability and detection accuracy for objects at different scales.Lastly,EIoU is utilized as the bounding box loss function to enhance the model's target localization accura-cy.Experimental results on the VOC dataset show that,compared with the YOLOv8n model,the precision,recall,and average precision mean of the BF-YOLOv8n model have been improved by 5.7,5.1 and 5.0 percentage points,respectively.Validation on the COCO dataset further demonstrates that the BF-YOLOv8n model has achieved enhancements in all performance metrics,fully testifying its excellent generalization and robustness.关键词
高级筛选特征提取/双向特征融合/YOLOv8n算法/多尺度特征/小目标检测头Key words
advanced screening feature extraction/bi-directional feature fusion/YOLOv8n algorithm/multi-scale features/small target detection head分类
信息技术与安全科学引用本文复制引用
雷亮,陈小庆,周华勇,徐山雯,陈毅,刘学涵,赵锦..基于多尺度特征融合的改进YOLOv8n车辆与行人目标检测算法[J].重庆科技大学学报（自然科学版）,2025,27(3):57-69,13.基金项目
2021年重庆市属本科高校与中国科学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）
重庆市教委科学技术研究项目"基于YOLO-Pose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）
