SEMACC: Efficient multistep attention crowd counting network based on shuffle module

Abstract

At present, crowd counting is a hot issue, although there have been many studies. However, most of the research is oriented toward improving detection accuracy, ignoring the constraints of monitoring computing power in practical application scenarios. In this paper, we give three challenges to crowd counting and propose a dense crowd counting network called SEMACC based on a deep neural network. SEMACC refers to the lightweight neural network ShuffleNetV2 to design the backbone network ShuffleUnit, and designs and proposes the lightweight feature extractor EMA module with multi-step attention integration centered on dilated separable convolution. The EMA module can effectively extract the mixed features of the channel, spatial and branch attention. Finally, the Decoder module is used to parse the feature into a crowd density map. The experimental results show that the parameters of SEMACC are only 2.85M, and the average accuracy is 91.3% in ShanghaiTech datasets, and satisfactory accuracy is obtained with less calculation.