基于通道域注意力机制的人群密度估计算法研究

Population density estimation algorithm based on channel domain attention mechanism

人群密度估计对于人流监控与人群安全具有重要意义。针对现有算法主要通过不同尺度感受野来学习特征,而无法高效利用多尺度特征的问题,文中设计了一个基于通道域注意力机制的特征融合模块。该模块可以在训练模型时学习特征融合的分布情况,以高效利用多尺度特征。此外,为解决人群数据集的样本有限问题,文中采用了多规模数据增广来训练模型。将新模型在Shanghaitech数据集上进行测试,并在陕西省某旅游景区人群计数数据集上进行验证。实验结果显示,基于通道域注意力机制的人群密度估计算法在MAE与MSE上均优于MCNN,证明了该方法在人群密度估计领域具有良好的应用价值。

Population density estimates are important for population flow monitoring and population safety in scenic spots.The existing algorithms mainly learn the features through different scales of receptive fields,and cannot utilize multi-scale features efficiently.Aiming at this problem,a feature fusion module based on the channel domain attention mechanism is designed.This module can learn the distribution of feature fusion when training the model to make efficient use of multi-scale features.In addition,in order to solve the problem that the sample of the crowd data set is very limited,multi-scale data augmentation is used to train the model.The new model was tested on the Shanghaitech dataset and validated on a population density estimation dataset from a tourist attraction in Shaanxi Province,China.The experimental results show that the population density estimation algorithm based on the channel domain attention mechanism is superior to MCNN on MAE and MSE.This method has a good application value in population density estimation.