Crowd and Group Detection in Shopping Centre Using Mask R-CNN

This paper presents a mask R-CNN model and a dataset to detect crowd and group of people in one of the shopping areas in Indonesia. The dataset presented is the real condition of people behavior during Covid-19 pandemic. The dataset created have a total of 760 Pictures with more than 4500 total annotated objects classified in three classes (Person, Group, and Crowd) that is divided into train, validation, and test dataset. The model then created using Mask R-CNN with Resnet 101 Backbone using COCO Pre-weight. Using a dynamic learning rate, the model has an accuracy of 88.35% and 0.772 validation-loss while tested on test dataset. Our contribution includes creating a workflow of mask R-CNN implementation that could be implemented in various shopping centers and a Crowd detection model that can be use to create an automatic monitoring system.