基于改进YOLOv3的人体目标检测算法研究  

Human Object Detection Algorithm Based on Improved YOLOv3

作　　者：梁芷 毋涛[1] 薛岩松 LIANG Zhi;WU Tao;XUE Yansong(School of Computer Science,Xi'an Polytechnic University,Xi'an 710600)

机构地区：[1]西安工程大学计算机科学学院,西安710600

出　　处：《计算机与数字工程》2024年第12期3503-3508,3554,共7页Computer & Digital Engineering

基　　金：陕西省科技成果转移与推广计划(编号:2019CGXNG-018)资助。

摘　　要：针对YOLOv3在人体检测上对小目标对象检测精度低、漏检率高,以及其检测速度无法满足现阶段实时检测需求的问题,论文提出基于改进YOLOv3的人体目标检测算法。首先,对VOC数据集中person类别使用K-means++聚类算法重新聚类,生成新的先验框,优化先验参数;其次,对YOLOv3骨干网络Darknet53进行通道剪枝,得到轻量化的骨干网络LR-Darknet41,减少模型参数,提高检测速度;最后,在特征融合部分,将部分深浅层特征通过RFB-s模块进行融合,扩大感受野,增强对小目标对象的检测。实验结果表明,所提出的改进算法相较原算法,漏检率降低3.7%,检测精度提高4.1%,检测速度达到53.6帧/秒。There are more problems that YOLOv3 have low detection accuracy,high missed detection rate for small target ob-jects in human detection,and YOLOv3's detection speed cannot meet the needs of real-time detection scene.To deal with these prob-lems,this paper proposes a human object detection algorithm based on improved YOLOv3.Firstly,the K-means++algorithm is uti-lized to cluster the target boundaries in the VOC data set of the person,and the priori parameters of the network are optimized by the clustering results.Secondly,the algorithm acquires a lightweight LR-Darknet41 by pruning the backbone network structure of Dark-net53,which can decrease the parameters of the model and improve the detection speed.Finally,the fusion of the shallow and deep features is achieved by using RFB-s,which can expand the receptive field and augment the detection of small-scale human target.The data show that,compared with the original algorithm,the improved algorithm reduces the missed detection rate by 3.7%and in-creases the detection accuracy by 4.1%,and the detection speed reaches 53.6 frames/s.

关 键 词：YOLOv3 人体检测 LR-Darknet41网络 特征融合 RFB-s 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]