Multi Scale Feature Fusion Crowd Density Estimation Technology Based on Residual Network

The multi-scale problem caused by the camera perspective is different from person to person. To achieve the objective, a multi array scale fusion crowd density estimation method based on residual network is offered in this research. The improved ResNet network is adopted in the front end of the network to extract the crowd features, and the output feature map is compressed into 1/8 of the original map to improve the accuracy of the prediction density map. The back-end uses three array dilated convolution modules with different dilated ratio to acquire the multi-scale features of the crowd, so that the network can obtain more scale details and edge information. At the end of the network,
$1\times 1$
convolution pairs are cascaded to get higher quality predictive density map. Test the network with datasets Shanghaitech and UCF_CC_50 respectively, and the performances are more excellent to the current mainstream crowd density estimation methods. The MAE value is 61.2% higher than that of MCNN network and 6.4% higher than that of SANET network.