基于STN-YOLOv8的小目标行人检测算法  

Small Target Pedestrian Detection Algorithm Based on STN-YOLOv8

作　　者：刘惠临[1] 吴宇浩 江宇 Liu Huilin;Wu Yuhao;Jiang Yu(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan,Anhui 232001,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《黑龙江工业学院学报(综合版)》2025年第2期107-113,共7页Journal of Heilongjiang University of Technology(Comprehensive Edition)

基　　金：安徽省高等学校科学研究项目(项目编号:2022AH040113);安徽理工大学研究生创新基金项目(项目编号:20223X2131)。

摘　　要：针对行人检测中出现的小尺度目标误检和漏检等现象,提出一种改进的STN-YOLOv8行人检测模型。在特征提取阶段,模型引入了SPD-Conv卷积,有效克服了传统步长卷积导致的特征信息损失,从而进一步提升了模型对小目标的检测效果。此外,在模型的颈部网络中,集成了三重特征编码模块,以增强特征的表示能力。定位损失函数方面,设计了一种融合NWD Loss的定位损失函数,有助于提高小目标行人的定位准确性。实验结果表明,与原YOLOv8模型相比,改进后的模型在mAP_(50)和mAP_(50:95)指标上分别提升了2.6%和2.4%,并且对小目标错检、漏检的问题有明显改善。An improved STN-YOLOv8 pedestrian detection model is proposed to address the issues of small-scale object false positives and false negatives in pedestrian detection.In the feature extraction stage,the model introduces SPD Conv convolution,which effectively overcomes the feature information loss caused by traditional stride convolution,thereby further improving the detection performance of the model for small targets.In addition,a triple feature encoding module is integrated into the neck network of the model to enhance the representation ability of features.In terms of positioning loss function,a positioning loss function that integrates NWD loss has been designed to improve the accuracy of small target pedestrian positioning.The experimental results show that compared with the original YOLOv8 model,the improved model has improved by 2.6%and 2.4%respectively in terms of mAP_(50)and mAP_(50:95)indicators,and has significantly improved the problems of small target false detection and missed detection.

关 键 词：行人检测 STN-YOLOv8 深度学习 SPD-Conv 损失函数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]