静态图像中采用混合卷积结构进行人群密度估计  

Crowd density estimation using hybrid convolution structure in static images

作　　者：范绿源 仝明磊 李敏 南昊 Fan Lyuyuan;Tong Minglei;Li Min;Nan Hao(School of Electronics&Information Engineering,Shanghai University of Electric Power,Shanghai 200090,China)

机构地区：[1]上海电力学院电子与信息工程学院,上海200090

出　　处：《计算机应用研究》2020年第3期906-909,931,共5页Application Research of Computers

基　　金：上海市自然科学基金资助项目(16ZR1413300)。

摘　　要：提出了一种混合卷积神经网络用于人群数量的感知计算,在高度密集的场景中可以准确地预测人群密度图。模型仅由两个部分组成:前端为扩张卷积神经网络提取二维特征;后端采用分数步长卷积神经网络降低下采样中的信息损失。为了验证和分析算法性能,模型设计基于当前较为流行的Shanghai Tech数据集,使用回归问题的评价指标,即平均绝对误差(MAE)和均方误差(MSE)作为评估算法性能的标准。在Shanghai Tech(MAE=100.8)、UCF_CC_50(MAE=305.3)与WorldExpo’10数据集上进行测试,实验表明模型在密集场景下较以往的方法有效降低了MAE和MSE,提高了密集人群计数的准确率。This paper developed a hybrid convolution neural network for perceptual crowd counting,which could accurately predict density maps in extremely crowded scenes.It consisted of merely two components:the front-end was a dilated convolutional neural network to extract two-dimensional features;the back-end deployed a fractionally stride convolution to lower the loss of image information caused by down-sampling.This paper designed the model structure based on the dataset Shanghai Tech,then in an attempt to acknowledge and analyze the performance of the algorithm,and afterwards made use of the evaluation indicators of the regression problem,the average absolute error(MAE)and the mean-square error(MSE)as the criteria.Additionally,testing the method on Shanghai Tech(MAE=100.8),UCF_CC_50(MAE=305.3)and WorldExpo’10 datasets while the experiment results reveal that the proposed model can effectively reduce MAE and MSE when compared with previous methods.

关 键 词：密集场景 扩张卷积 分数步长卷积 密度估计 人群计数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]