Contagem de Fluxo de Pessoas Utilizando Aprendizado Profundo

Abstract

The customer's flow count counts the number of people entering the stores. This count allows you to extract different business metrics such as conversion rate of marketing actions, time of the visit, and people's traffic. The main objective of this dissertation is to propose, develop and evaluate a solution to count customers using security cameras. The proposal is to combine deep learning algorithms for counting people. Additionally, identify people who should not be counted, such as employees and collaborators. We collected and labeled videos in two different places. We used the labeled data to train the Yolov5 to define the count by People and RetinaFace by Face. The counting performed by the proposal was compared with the manual counting using a significance test. According to the test, there was no significant difference between the system predicted and the ground truth scores. The economic feasibility presented a cost of 24.4USD per month, considering 10 hours of video per day for cloud processing. The proposed solution does not require specific hardware and modifications in the store's spaces, being a promising alternative for this problem.