IMPLEMENTATION OF CROWD DETECTION MODEL IN COVID-19 PANDEMIC BASED ON CONVOLUTIONAL NEURAL NETWORKS USING UAV

• Main Author: Matheus, Leonard
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/72150
• Institution: Institut Teknologi Bandung

Social distancing is one of the solutions to break the transmission chain of COVID-19. However, human crowds are the main problem of close contact between humans who are close together. The human crowd detection model is needed that can estimate the distance between two or more people to prevent violations of social distance with a safe limit of less than 1.5 meters from the point of view of the UAV camera. CNN model training was performed on 9,600 images of humans, cyclists, and motorcyclists. The pre-trained model used is Single Shot Detection with the MobileNet, ResNet50, and ResNet101 architectures. In addition, the social distance estimation uses Euclidian distance with the average Indonesian human height as a reference, which is 1.6 meters. Social distance calibration is also carried out using the principle of projection from the different viewpoints of the UAV camera while flying. Based on the test results, MobileNet V2 was chosen as the crowd detection model. MobileNet V2 has a model size below 100 Megabytes and the average detection runtime for a single image is only 0.606 seconds, so that it fits the UAV companion computer load that can handle. MobileNet V2 is also able to detect crowds of people well, as evidenced by a precision value of up to 84.9% (IoU=0.50:0.95) and a sensitivity (recall) value of up to 87.8% (MaxDets=100). Finally, a program was successfully developed to calculate violations using social distance estimation.