Dense Crowds Detection and Surveillance with Drones using Density Maps

Detecting and Counting people in a human crowd from a moving drone present challenging problems that arise from the constant changing in the image perspective and camera angle. In this paper, we test two different state-of-the-art approaches, density map generation with VGG19 trained with the Bayes loss function and detect-then-count with Faster R-CNN with ResNet50-FPN as backbone, in order to compare their accuracy at counting and detecting people in different scenarios taken from a drone in flight. We show empirically that both proposed methodologies perform well for detecting and counting people in sparse crowds when the drone is near the ground. Nevertheless, Bayes Loss provides better accuracy on both tasks while also being lighter than Faster R-CNN. Furthermore, Bayes Loss outperforms Faster R-CNN when dealing with dense crowds, proving to be more robust to scale variations and strong occlusions, hence being more suitable for surveillance applications using drones.