Crowd Density Estimation Based on Multi-Column Hybrid Convolutional Network

This paper studies an accurate counting model for dealing with highly crowded people, multi-column hybrid convolutional neural network model. The model is mainly composed of three parts. The first part uses the first ten layers of VGG-16 convolutional network for image feature extraction. The middle layer is a dilated convolution with three rows of “jaggy” dilation rates, and each row uses the Resnet-block connection method, which is used primarily to perceive human head features of different sizes. Compared with a variety of image up-sampling ways, in the third part of the model, this paper tries to use a combination of bilinear interpolation and convolution to up-sample image features, and research shows that this method effectively reduces the model error. In this experiment, the average absolute error (MAE), mean square error (MSE) and average relative error (MRE) are used as evaluation indicators, and experiments on the ShanghaiTech dataset proves that the network works well.