PDDNet: lightweight congested crowd counting via pyramid depth-wise dilated convolution

The accuracy of crowd counting is susceptible to scale variations of crowd head in the congested scene. Some counting networks, such as crowd density pre-classification networks or multi-column counting networks, are proposed to model the different scales of crowd head. However, most of them own a complex network structure with many network parameters, making deploying a crowd counting network in practical application scenarios challenging. To this end, we propose a lightweight crowd counting network termed PDDNet. The front-end of the PDDNet chooses the first 13 layers of GhostNet to extract the crowd feature, and the back-end of the PDDNet is implemented with the proposed lightweight pyramidal convolution modules (LPC) to extract the multi-scale features. Finally, the extracted multi-scale features are fed to transposed convolution layers to regress the final crowd density map. We conduct extensive experiments on the commonly-used crowd counting datasets, i.e., ShanghaiTech, UCF_QNRF, and NWPU_Crowd. The experiment results show the superiority of our model compared with state-of-the-art methods.

Automated summary

In this study, we propose a lightweight crowd counting network called PDDNet, which utilizes GhostNet to extract crowd features at the front-end and employs lightweight pyramidal convolution modules to extract multi-scale features at the back-end. The extracted features are then fed to transposed convolution layers to regress the crowd density map. Extensive experiments on commonly-used crowd counting datasets demonstrate the superiority of our model compared to state-of-the-art methods.