【摘要】 为解决密集人群计数任务中多列卷积核独立训练的限制及缺少针对性优化的问题,提出融合多尺度特征的密集人群计数算法SGNet。通过设计一种围绕相同感受野SRF(samereceptivefield)的特征融合方法,达到强化不同特征列之间的关联性,获得更多的特征细节和特征信息的目的;融合网格赢家通吃GWTA(gridwinner-take-all)的思想设计损失函数,通过计算区域损失值着重优化重要特征。实验结果表明,与基线模型相比SGNet在任一数据集上的检测效果均有一定程度的提升,验证了该模型具有较强的鲁棒性及可移植性。更多还原

【Abstract】 To solve the limitation of independent training of multi column convolution kernel and the lack of targeted optimization in dense crowd counting task,a dense crowd counting algorithm SGNet integrating multi-scale features was proposed.By designing a feature fusion method around the same receptive field SRF(same receptive field),the correlation between different feature columns was strengthened and more feature details and feature information were obtained.The loss function was designed by integrating the idea of GWTA(grid winner-take-all),and the important features were optimized by calculating the regional loss value.Experimental results show that,compared with the baseline model,the detection effect of SGNet on any dataset is improved to a certain extent,and it is verified that the model has strong robustness and portability.更多还原