Research on Crowd Counting Based on Attention Mechanism and Dilation Convolution

Abstract

Aiming at the problem of low accuracy of population counting in existing algorithms, we propose a density estimation algorithm combining attention mechanism and dilated convolution. In this paper, the basic framework of feature extraction consists of part of the network layer of VGG-16 and attention mechanism. Then, replace part of the pooling layer and fully connected layer of the original network with a zigzag dilation convolution module to effectively compensate for the information loss caused by the pooling layer. Specially, the ability of the network model to extract features is improved by fusing the feature information of the high and low layers, thereby improving the counting performance of the model. We compare our method with the other state-of-the-art works, and the experiment results demonstrate the superiority of our method, which shows that the proposed method has high accuracy, strong adaptability and robustness, and can well adapt to the detection of people of different densities.