SC2Net: Scale-aware Crowd Counting Network with Pyramid Dilated Convolution

Accurate crowd counting is still challenging due to the variations of crowd heads. Most of crowd counting methods adopt multi-branch networks to extract multi-scale information. However, these networks are too complex to be optimized. To solve these problems, we propose an efficient scale-aware crow...

Full description

Saved in:
Bibliographic Details
Published in:Applied intelligence (Dordrecht, Netherlands) Netherlands), 2023-03, Vol.53 (5), p.5146-5159
Main Authors: Liang, Lanjun, Zhao, Huailin, Zhou, Fangbo, Zhang, Qing, Song, Zhili, Shi, Qingxuan
Format: Article
Language:English
Subjects:
Artificial Intelligence
Coders
Computer Science
Convolution
Encoders-Decoders
Feature extraction
Feature maps
Localization
Machines
Manufacturing
Mechanical Engineering
Modules
Processes
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Accurate crowd counting is still challenging due to the variations of crowd heads. Most of crowd counting methods adopt multi-branch networks to extract multi-scale information. However, these networks are too complex to be optimized. To solve these problems, we propose an efficient scale-aware crowd counting network named SC2Net, which adopts the encoder-decoder framework. The encoder uses the first ten layers of VGG16 to extract the primary feature information. The decoder is mainly consisted of our proposed residual pyramid dilated convolution (ResPyDConv) modules to regress predicted density maps. Specifically, the ResPyDConv module is composed of pyramid dilated convolution (PyDConv). Each PyDConv adopts dilated convolutions with different dilated rates. PyDConv divides feature maps into different groups and extracts multi-scale feature information. Extensive experiments are conducted on ShanghaiTech, UCF_CC_50, UCF_QNRF, and NWPU_Crowd datasets. Qualitative and quantitive results show the superiority of our proposed network to the other state-of-the-art methods.
ISSN:0924-669X
1573-7497
DOI:10.1007/s10489-022-03648-4