摘要
随着城镇化进程的加速,高密度人群场景的监控和分析需求激增。基于弱监督学习的人群计数方法仅使用计数级标签,近年来受到研究人员越来越多的关注。目前,弱监督人群计数仍然面临着人群尺度差异和复杂背景干扰的挑战。为解决这些问题,提出一种细节增强引导的弱监督人群计数方法。该方法在特征提取阶段分为2个分支。主干分支选用ResNet网络和Transformer的融合网络,进行局部特征和全局特征的提取;辅助分支设计了细节增强模块,旨在引导和强化图像中的细节信息。随后,通过注意力加权融合模块将2个分支输出的特征进行有效融合,融合后的特征进入计数回归模块进行人群计数。在多个标准人群计数数据集上进行了试验,试验结果表明该方法可以达到高精度的人群计数。
With the acceleration of the urbanization,there has been a surge in demand for monitoring and analysis of high-density crowd scenarios.Crowd counting methods based on weakly-supervised learning,which only rely on count-level labels,have attracted increasing attention in recent years.At present,weakly-supervised crowd counting still faces the challenges of crowd scale differences and complex background interference.In order to solve these problems,we propose a weakly-supervised crowd counting method guided by detail enhancement.Specifically,the method is divided into two branches in the feature extraction stage.The main branch integrates ResNet network with Transformer to extract local features and global features.The auxiliary branch includes a detail-enhanced module,which aims to guide and strengthen the detail information in the image.Then,the features extracted by the two branches are effectively integrated by the adaptive fusion module,then the integrated features are passed into the counting regression module for crowd counting.Experiments on several standard crowd counting datasets are also conducted,and the results show that the proposed method achieves excellent counting performance.
作者
张德
蔡雨航
ZHANG De;CAI Yuhang(School of Intelligence Science and Technology,Beijing University of Civil Engineering and Architecture,Beijing 100044;Beijing Key Laboratory of Super Intelligent Technology for Urban Architecture,Beijing University of Civil Engineering and Architecture,Beijing 100044)
出处
《北京建筑大学学报》
2026年第1期137-148,共12页
Journal of Beijing University of Civil Engineering and Architecture
基金
国家自然科学基金项目(62271035)
北京市自然科学基金项目(4232021)。
