Real-Time On-Board Counting System in Crowded Scenes

Abstract

ABSTRACT : In public spaces, crowd counting is a useful tool for situational awareness. Interest in computer vision has grown significantly as a result of the fascinating yet challenging problem of automated crowd counting using images and videos. During the past few decades, many approaches were proposed to count people in crowded scenes and which have made a remarkable progress. However, there is not a crowd counting model that works for all applications (e.g., non real-time, real-time), and which can be deployed in different hardware platforms with different computational capabilities (e.g., cameras, drones, servers, edge devices, mobile phones etc.). This thesis aims to give the background of crowd counting, and explains our realtime crowd counting and face detector model called YOLO-CROWD, which solves the problem of occlusion in YOLO models and which can be deployed on edge devices. Additionally, this thesis describes our platform CrowdCounting.AI, which implements dense and very computationally-demanding models that run on powerful devices for the best accuracy, and which also implements our model YOLO-CROWD for realtime applications.