Crowd counting method based on deep residual network

Abstract

The invention discloses a crowd counting method based on a deep residual network. The method applies the deep residual network to extract the characteristic of each frame of image in a crowd monitoring video, wherein the input of the deep residual network is one frame of image; through 5*5 kernel convolution and pooling, an initial characteristic graph is obtained; through ten residual network units, characteristics are extracted; a main branch obtains a crowd density graph corresponding to an input image through 1*1 kernel convolution; an auxiliary branch obtains a people number corresponding to the input image through the 1*1 kernel convolution; and finally, through the integration of the crowd density graph, the people number estimation value of the input image is obtained. Each residual network unit has the structure that a 3*3 conventional kernel is accessed after a 1*1 convolution kernel, then, the 1*1 convolution kernel is accessed, a batch normalization operation and a linear rectification operation are added after each convolution kernel, and meanwhile, the output of a previous residual network unit also serves as the input of a next residual network unit through the 1*1 kernel convolution. By use of the method, an influence on crowd counting by scene transformation can be reduced, and a stable crowd counting effect is obtained.

Description

technical field [0001] The invention relates to a crowd counting method in a surveillance video, in particular to a crowd counting method based on a deep residual network. Background technique [0002] The current mainstream crowd counting methods mainly include the quantitative regression algorithm based on foreground features and the density map regression algorithm based on neural networks. The main disadvantage of the former is that feature extraction depends on the foreground segmentation effect of video images, and the trained model is limited. Specific scenarios; the main disadvantage of the latter is that it needs to use the sub-network structure to achieve multi-scale feature extraction, the scale jumps are large, and the obtained model parameters are also large, which is not suitable for current monitoring equipment with low computing power. Contents of the invention [0003] The purpose of the present invention is to provide a crowd counting method based on a de...

AI Technical Summary

Problems solved by technology

[0002] The current mainstream crowd counting methods mainly include the quantitative regression algorithm based on foreground features and the density map regression algorithm based on neural networks. The main disadvantage of the former is that feature extraction depends on the foreground segmentation effect of video images, and the trained model is limited. Specific scenarios; the main disadvantage of the latter is that it needs to use the sub-network structure to achieve multi-scale feature extraction, the scale jumps are large, and the obtained model parameters are also large, which is not suitable for current monitoring equipment with low computing power.

Images