Deep Classification Technique for Density Counting

Abstract

Crowd counting, resulted from extensive analysis, is reflected by many aspects such as appearance similarity between people, background components and the inter-blocking in intense crowds. Current research is challenging these aspects by applying different types of architectures. In this paper, we propose a single conventional neural network for density counting based on four conventional layers. A comparison of our proposed network with Switched Conventional Neural Networks (Switch-CNN) approaches has been performed in order to evaluate its performance in terms of accuracy and loss. As a result, several experiments prove the effectiveness and efficiency of the proposed method. We got 94.6% and 0.2625 for both accuracy and loss respectively.