Single image crowd counting using deep learning models

Abstract

This MTech. thesis presents multiple approaches for crowd counting in single images using deep learning models. The research investigates multiple techniques to improve the accuracy of crowd counting, with a particular focus on high-density crowd images where existing models often yield suboptimal results. The frequency domain approach was employed to provide compact and effective supervision for the network. By transforming the density map into the frequency domain using Fast Fourier Transform (FFT), global spatial information was incorporated without relying on external algorithms. This approach enabled the network to leverage frequency-based representations for crowd counting.