Journal of Applied Science and Engineering (Nov 2022)
Crowd Density Estimation Based On Multi-scale Information Fusion And Matching Network In Scenic Spots
Abstract
The crowd density estimation has important application value in intelligent safety prevention, traffic safety and tourist attractions safety prevention, etc. The task of crowd density estimation is to estimate the crowd density distribution by extracting and analyzing the crowd features. Traditional crowd density estimation methods show great differences in perspective changes in 2D images, resulting in loss of feature spatial information and difficulty in extracting scale features and crowd features. In this paper, we propose a novel crowd density estimation method based on multi-scale information fusion and matching network in scenic spots. A multi-scale feature extraction module is introduced to extract multi-scale features of different convolutional layers in matching networks. Through the combination of multi-scale asymmetric convolution and dilated convolution with different expansion rates, the expression ability of extracted semantic information and scale information is enhanced. Finally, in the multi-scale information fusion network, the semantic embedding method is used to introduce spatial information into high-level semantic information. The high level semantic information is introduced into the low level spatial information to enhance the feature expression. And the scale information is integrated with the spatial global context information to obtain the high quality density map and predict the crowd more accurately. Then, we conduct experiments on open data sets, and the results show that the presented model in this paper has good adaptability in scenic spots with large differences in crowd distribution, the average MSE with proposed method is below 15, which is the lowest value and it can extract features from different scenes to estimate density distribution and accurately count crowds.
Keywords
