摘要
为解决人群计数任务中因人群大小不一导致的计数偏差问题,提出一种基于特征融合的编解码卷积神经网络模型(CFFNet)。前端网络模块对输入的人群图像自动编码,提取不同尺度的人群特征语义信息;后端网络模块对编码后的人群特征信息进行融合和解码,得到最终的估计密度图。将该模型在4个公开数据集上进行实验,并与历年的主要方法进行对比,实验结果表明,该模型在ShanghaiTech PartA、UCSD和Mall数据集上取得了更好的MAE指标,优于目前的这些算法,验证了模型对不同的人群尺度具有很好的适应性。
To deal with the problem of counting bias caused by different crowd scales in crowd counting tasks,a codec convolutional neural network model(codec-on-feature-fusion-network)based on feature fusion was proposed.The front-end network module automatically encoded the input crowd images and extracted the semantic information of crowd features at different scales.The back-end network module fused and decoded the encoded crowd feature information to obtain the final estimated density map.The model was experimented on four publicly available datasets and compared with the main methods over the years.Experimental results show that the model achieves better MAE metrics on the ShanghaiTech PartA,UCSD and Mall datasets and outperforms these current algorithms,further verifying that the model has good adaptability to different crowd scales.
作者
邹敏
黄虹
杜渂
黄继风
ZOU Min;HUANG Hong;DU Wen;HUANG Ji-feng(College of Information and Mechanical and Electrical Engineering,Shanghai Normal University,Shanghai 200234,China;School of Finance and Business,Shanghai Normal University,Shanghai 200234,China;DS Information Technology Limited Company Shanghai,Shanghai 200032,China)
出处
《计算机工程与设计》
北大核心
2023年第7期2110-2117,共8页
Computer Engineering and Design
基金
上海市地方能力建设基金项目(19070502900)。
关键词
人群计数
卷积神经网络
编码器
人群尺度
解码器
特征融合
密度图
crowd counting
convolutional neural network
encoder
crowd scale
decoder
feature fusion
density map
