Abstract: The crowd counting algorithm of adversarial neural network based on multi-scale feature fusion is proposed to solve the difficulty in crowd feature extraction and information loss in feature fusion. First, shallow crowd features at different scales are extracted through multi-scale feature extraction structure with different size convolution kernels extracts. Next, shallow crowd features and deep-level crowd features of the convolutional network are connected through residual structure to achieve the fusion of crowd features at different scales and depths.Finally, the alternate learning between generator network and discriminator network is realized through adversarial method in order to generate high-quality crowd density maps. Experimental results on two datasets of ShanghaiTech and UCF_CC_50 show that this model can greatly improve the accuracy and robustness of crowd counting in complex crowd environments compared with traditional neural network models.

Key words: crowd counting, adversarial neural network, feature fusion, complex crowd environment