Deep Learning for Crowd Image Classification for Images Captured Under Varying Climatic and Lighting Condition

Abstract

Most of recent events have attracted a lot of attention towards importance of automatic crowd classification and management. COVID-19 is the most setback for the entire world. During these events proper breakout and public crowd management leads to the requirement of managing, counting, securing as well as tracking the crowd. But automatic analysis of the crowd is very challenging task because of varying climatic and lighting conditions, varying postures etc. During this paper we have developed PYTHON based system for automatic crowd images classification using Deep learning. This paper is the first attempt for automatic classification of crowd images. We have prepared the dataset of crowd classification consisting of three categories. The proposed methodology of crowd classification starts with preprocessing during which we have used median filtering for noise removal. Deep learning models are developed using 70% training images. The performance of the system is evaluated for various deep learning algorithms including one block VGG, two block VGG and three block VGG. We have also evaluated the performance of three block VGG using dropout. VGG16 transfer learning based crowd classification is developed using PYTHON. Using VGG16 transfer learning we achieved the accuracy of 69.44.% which is highest among all deep learning classification models during this study