A novel multi-scale crowd counting algorithm using python (PyMCCA)

Abstract

This research was conducted to create a Multi-Scale Crowd Counting Algorithm. This algorithm implements a patch-based inference that is able to give the number of people in an image whether the image contain a small or large crowd size. The algorithm was designed to perform within the standard of the existing small scale and large- scale crowd counting algorithms using Convolutional Neural Network Architectures such as Inception V3 for the crowd classifier, Resnet101 for the small-scale model, and VGG-16 for the large-scale model. This algorithm was implemented along with a user interface using PyQt5 GUI designer to make it more convenient to use. The algorithm works by identifying whether an image contains a small or large crowd, then implements a model best suited for the image. If the image is determined to contain small amount crowd, the image will be divided into 4 patches then a Faster R-CNN model trained on human head and body annotations will be implemented, while if the image contains a large number of crowd, the image will be divide into 9 patches then an inference algorithm based on CSR model will be implemented.