RDNet: A reinforced-deformable convolutional network for crowd counting

Crowd counting is still a challenging task in crowded scenario due to heavy occlusions, appearance variations and perspective distortions. Thanks to the development of deep convolutional neural networks, especially the deformable convo-lution neurral networks, great progress has been made for crowd counting. However, existing deformable convolution networks cannot get satisfactory performance due to limited receptive field that is not conducive to crowded object detection. In this paper, a novel end-to-end Reinforced-Deformable convolutional neural Network (RDNet) is proposed. First, a multi-scale context information fusion strategy with skipping connection and shortcut connection is utilized to obtain rich semantic features and local details of the crowd distribution. Then a reinforced-deformable convolution module is designed as the back-end of the network, which can sample with adaptive offsets and scalars to enlarge the receptive field for more accurate edge information and crowd distribution localization. Finally, an impulse function convolved with cuphead Gaussian kernel is applied to generate density map for counting. Extensive experimental results on four popular datasets, namely, Shangha-itech, UCF _CC_50, WorldEXPO'10, and UCF-QNRF, show that our proposed approach outperforms other state-of-the-art methods.