Person Counting Based On Graph Relation Network Using UAV

Abstract

Person counting at the construction site is of great significance to person scheduling and progress supervision. This work aims to develop a person counting framework, which is composed of three modules: person detection network, person re-identification module and person counting module. In order to solve the problem of repeated counting caused by Unmanned Aerial Vehicles (UAVs) taking multiple shots of the same person, we propose a Graph Similarity-based Person Counting Network (GSPCN), which can re-identify persons and avoid counting the same person multiple times. We tested the proposed framework both on simulation environment and real datasets (our data was collected from multiple construction sites in Shenzhen), It is empirically superior to the most advanced methods available.