Density Map Based Estimation of Crowd Counting Using Vgg-16 Neural Network
G Anusiya, Nandha Kumar S, Siva Sangari V, S Santhi · 2023
The system is to evolve a method that can precisely evaluate the count of crowds from a raw image obtained with random crowd size and perspective. In the current world scenario, in many places, they are using the old traditional ways, like preserving registers and using sensors, in the entrance for counting the number of people in the crowd. These approaches are ineffective. Three methods namely counting by Regression, Detection and Density Estimation can be used to perform crowd counting. In our approach, we are counting by Density Map based Estimation. It is a VGG-16- based Convolutional Neural Network that is trained to categorize the input images into various density classes. We are using an architecture that contains multi-column based CNN, each column CNN learns can adapt to changes in people or heads sizes caused by perspective effects or differences in the image by using filters based on the sizes. It aims to create high-dimensional maps from the given image which can be combined with the provisional data obtained from the density classes. Additionally, a real density map is precisely calculated using different kernels, which don't require any knowledge about the input image's outlook map. In this case, a Gaussian filter was used to create the ground truth density map, and when it was submitted to the neural network, it created the equivalent density map. We have collected a dataset that contains 1199 photos with over 2, 20,000 heads annotated because the existing dataset does not capture all the circumstances. We conduct experiments on the collected dataset as well as the old dataset in order to validate the current model and method used. In particular, our strategy beats all currently used methods for the proposed simple MCNN model.