基于多任务学习及由粗到精的卷积神经网络人群计数模型

Crowd Counting Model of Convolutional Neural Network Based on Multi-task Learning and Coarse to Fine

人群计数是指计算单张图像或单个视频帧中人的数目,为了解决人群任务的计数不够准确的问题,提出了一种基于多任务学习及由粗到精的卷积神经网络人群计数模型。首先,多任务学习是指引入与原始任务相关的辅助任务,指导主要任务的学习,人群密度估计是人群计数模型的主要任务,人群分割任务作为辅助任务以提高网络性能。其次,由粗到精策略表明人群计数模型预测密度图是一个由粗糙到精细的过程,即生成粗糙且不准确的人群密度图,结合人群分割图后得到准确的人群密度图。在Shanghai Tech数据集Part A部分、Part B部分和UCF_CC_50数据集上的实验表明,所提人群计数模型相比之前最好的CSRNet模型绝对误差分别降低了4.55%,14.15%,19.09%,均方误差分别降低了10.00%,19.09%,19.47%,显著提高了人群计数模型的准确性和鲁棒性。

Crowd counting refers to counting the number of people in a single image or a single video frame.In order to solve the problem of insufficient counting of crowd tasks,a crowd counting model based on multi-task learning and coarse to fine convolutional neural network is proposed.Firstly,multi-task learning means introducing auxiliary tasks related to the original task to guide the learning of the main tasks.The crowd density estimation is the main task of the crowd counting model,and the crowd segmentation task is used as an auxiliary task to improve network performance.Secondly,the proposed crowd counting model is able to predict the density map from coarse to fine.A rough and inaccurate crowd density map is generated,which is combined with the crowd segmentation map to obtain an accurate crowd density map.Experiments on the Shanghai Tech dataset Part A and Part B,and UCF_CC_50 dataset show that the proposed crowd counting model outperforms the state of the art CSRNet models by 4.55%,14.15%and 19.09%respectively,and the mean square error is reduced by 10.00%,19.09%and 19.47%respectively compared with the SOTAs.The proposed model significantly improves the accuracy and robustness of the crowd counting model.