Abstract

It has been almost a year into the COVID-19 pandemic, and the Philippines has been adapting to the necessary changes in lifestyle according to the global health guidelines. One such regulation that many businesses are relaxing is in the social distancing between individuals. In the transition to the new normal, how can we effectively and non-obstructively enforce social distancing?

To solve this, two crowd counting and one crowd detection model was applied on an image dataset to count and detect pedestrians, respectively. The two crowd counting models were a custom and a MobileNetV2 architecture designed to count the pedestrians in the image and obtain a Mean Absolute Error of 1.87 and 2.47, respectively. The crowd detection model uses the MobileNetV2-YoloV3-Lite as its base and utilizes a Homography Transform to approximate the pedestrians’ locations and the distances between them despite the depth of field. The crowd counting and crowd detection models provide an indirect and direct method to analyze if social distancing is being abided by within the image.