Context-aware pyramid attention network for crowd counting

Achieving accurate crowd counting still faces many challenges due to continuous scale variations. To this end, we present an innovative Context-Aware Pyramid Attention Network for crowd counting which is realized by extracting rich contextual features and dealing with dependencies on space and channels. To extract rich contextual features, we propose a context-aware pyramid feature extraction module. With this module, the rich contextual feature extraction is implemented by dividing the input features into four blocks with different scales. In addition, we design an attention module consisting of a space attention block and a channel attention block which are responsible for dealing with the interdependence on feature information in the spatial dimension and the channel dimension, respectively. We first perform an ablation study to evaluate the effectiveness of each component on ShanghaiTech Part_A. Then we conduct several comparative experiments with several state-of-art methods on five challenging crowd counting datasets, including the ShanghaiTech, the UCF_CC_50, the UCF-QNRF, the WorldExpo’10, and NWPU-Crowd datasets. Experimental results demonstrate that the mean absolute error and mean square error of CAPAN are significantly reduced compared to the state-of-art methods in counting accuracy, and at the same time, the quality of density maps and the false recognition rate are improved. These results validate the effectiveness of CAPAN in crowd counting.