Crowd Counting Method Based on Deep Neural Network

Abstract

With the progress of society and the rapid development of economy, the pace of urbanization construction is promoted. The number of people flocking to cities increases sharply, and the population density keeps increasing. Therefore, crowd counting has important application value in the field of intelligent security. Aiming at the problems of background interference, serious occlusion between crowds, insufficient information of context feature extraction and huge scale change in Crowd Counting and Crowd density estimation tasks. The CAI-CAN model proposed in this paper is composed of four parts, which are front-end network, CAI Module, CA Module and back-end network. The input image is transmitted to the front-end network for primary feature extraction. CAN module performs multi-scale feature fusion on the image after feature extraction. CA module consists of two parts: coordinate information embedding and coordinate attention generation. Given the input, use two pooling kernels to encode along the horizontal and vertical coordinates. Finally, the processed features are sent to the back-end network of density generation module, and the density map is obtained through convolution. Our method is validated on Shanghaitech dataset, UCF-CC_50 dataset and JHU-Crowd ++ dataset. Experimental results show that compared with CAN and other methods, the proposed method achieves better accuracy and robustness.