摘要
经典的人群计数都是针对一些高清图片的全局检测,效果提升显著,但真实的站台数据为低分辨率图片,并且存在遮挡和站外目标等干扰因素,原始的检测方法不能很好地适应它.因此本文提出了一种基于YOLOv5的站台人群区域密度分析模型,从而有效帮助公交根据站台密度进行合理调度.为获取更多特征信息,本文使用SPD卷积模块替换了原始的池化层,并创新地结合了CBAM注意力模块和区域检测模块,使得优化后的模型能够针对站台中的特定区域进行密度分析.本文还针对不同站台的特定区域提出了区域切换算法,从而使模型的区域分析更加灵活.与传统的YOLOv5相比,优化后的模型在测试集上的mAP提升了1.1%,最终的检测指标提升了18.2%.因此,本文提出的模型更适用于站台数据的密度分析.
The classical crowd counting methods are global detection for high-definition images,with significant improvement in performance.However,the actual platform data is low resolution images and there are interference factors such as occlusion and off-site targets,the original detection methods cannot adapt well to this dataset.Therefore,this article proposes a platform crowd area density analysis model based on YOLOv5,which effectively helps public transportation schedule reasonably based on platform density.To obtain more features,this article replaces original pooling layers with SPD convolution module,and innovatively combines the CBAM attention module and region detection module,enabling the optimized model to perform density analysis for specific areas in the platform.This article also proposes a region switching algorithm for specific areas of different platforms,making the region analysis of the model more flexible.Compared with the traditional YOLOv5,the optimized model improved mAP by 1.1%on the test set and the final detection index by 18.2%.Therefore,the model proposed in this article is more suitable for density analysis of platform data.
作者
李钊宇
陈庆奎
LI Zhaoyu;CHEN Qingkui(College of Optoelectronic Information and Computer Engineering,University of Shanghai for Science and Technology,Shanghai 200093,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2024年第11期2732-2738,共7页
Journal of Chinese Computer Systems
基金
国家自然科学基金项目(61572325)资助.
作者简介
李钊宇,男,1999年生,硕士研究生,CCF会员,研究方向为计算机视觉、目标检测,E-mail:lys063311@163.com;通信作者:陈庆奎,男,1966年生,博士,教授,博士生导师,CCF会员,研究方向为计算机集群、并行计算、人工智能等.
