基于深度学习的密集行人检测场景算法研究
Algorithm Research on Dense Pedestrian Detection Scene Based on Deep Learning
马明杰 1马小陆 1唐得志 1赵远 1齐晶晶 2瞿元3
作者信息
- 1. 安徽工业大学 电气与信息工程学院,安徽 马鞍山 243002;安徽工业大学 芜湖技术创新研究院,安徽 芜湖 241002
- 2. 安徽达尔智能控制系统股份有限公司,安徽 芜湖 241002
- 3. 奇瑞汽车股份有限公司,安徽 芜湖 241006
- 折叠
摘要
针对密集行人检测场景存在目标尺度过小以及目标遮挡等问题,提出一种基于改进YOLOv7 的密集行人检测算法.首先在特征提取网络引入MobileNet注意力模块,减少模型计算量和增强特征提取能力;其次在特征融合网络加入BepC3 模块,提升了行人多尺度特征融合的能力;最后采用WD-Loss作为定位损失函数,提高模型检测的定位精度.在Wider-Person拥挤行人检测数据集上进行训练和验证,实验结果表明改进后的算法模型AP 50 精度达到了 0.784,领先原YOLOv7 算法0.031.
Abstract
A dense pedestrian detection algorithm based on the improved YOLOv7 is proposed to ad-dress the issues of small target scale and occlusion in dense pedestrian detection scenarios.Firstly,the MobileNet attention module is introduced into the feature extraction network to reduce the model compu-tation and enhance feature extraction capabilities.Secondly,the addition of the BepC3 module in the feature fusion network enhances the ability of pedestrian multi-scale feature fusion.Finally,WD-Loss is used as the localization loss function to improve the localization accuracy of the model detection.Trai-ning and validation were conducted on the Wider-Person crowded pedestrian detection dataset,and the experimental results showed that the improved algorithm model AP50 achieved an accuracy of 0.784,leading the original YOLOv7 algorithm by 0.031.
关键词
密集行人检测/YOLOv7/MobileNet/BepC3/WD-LossKey words
dense pedestrian detection/YOLOv7/MobileNet/BepC3/WD-Loss引用本文复制引用
基金项目
国家自然科学基金资助项目(62172004)
国家自然科学基金资助项目(61872004)
安徽省科技重大专项(202003a05020028)
安徽省高校协同创新项目(GXXT-2023-020)
芜湖市核心技术攻关科技计划项目(2022hg10)
芜湖市科技计划项目(2023kx17)
出版年
2024
NETL
NSTL
万方数据