Lightweight solution to background noise in crowd counting

Abstract

This paper proposed Dilated Compact Convolutional Neural Network (DCCNN) for single-image crowd density estimation from the original lightweight C-CNN. DCCNN is an enhancement of lightweight C-CNN compensated for lack of mechanisms to alleviate background noise using dilated convolution and average pooling. The performance of our proposed model improves significantly on medium and spared crowd scenes in ShanghaiTech part B dataset, achieving 18% lower MAE compared to C-CNN while requiring virtually no additional computational costs.