The recent introduction of drones in a wide range of applications such as visual surveillance, rescue, and entertainment, is accompanied by the demand of their safe operation. Apart from the issue of making drones robust to failures (electrical, mechanical etc.), an important step towards increased safety constitutes in defining no-fly zones above or near crowds, since a drone may operate close to crowds, and is potentially exposed to environmental hazards, unpredictable failures or pilot errors that may lead to accidents. Thus, it is of utmost importance for the drone to be able to detect crowds in order to define no-fly zones and proceed to re-planning during flight.

Example of crowd detection  – the image on the right shows the heat-map created by the software

The problem of crowd detection refers to the process of automatically determining the presence of a mass of people in the same location. Given an image it should be determined if there is crowd and where it is. In general, crowd detection is a prerequisite component for a wide spectrum of applications, including crowd management, abnormal human behaviour recognition during crowded events such as marathons, demonstrations and music concerts, crowd motion analysis, etc. However, the visual ambiguities stemming from the small object sizes and the variations of viewing angle, crowd density, scale etc. render crowd detection in aerial images captured from drones a challenging task.

Example of crowd detection at a marathon: Group of runners on open street

Towards this end, a crowd detection method for drone flight safety, using fully Convolutional Neural Networks (CNN) has been developed by the MultiDrone partner, Aristotle University of Thessaloniki. The focus is on providing a lightweight CNN model, in order to comply with the low computational cost requirements of the specific application (since the algorithm shall run on the drone computing platform), that can distinguish between crowded and non-crowded scenes, captured from the drone camera. The output of the algorithm is in the form of heat-maps that represent the estimated probability of crowd existence in each location within the captured scene.

Example of crowd detection: The heat-map shows how the image is perceived differently than the previous one

More details can be found in the upcoming conference publication:

Maria Tzelepi, Anastasios Tefas, “Human Crowd Detection for Drone Flight Safety Using Convolutional Neural Networks”, in 25th European Signal Processing Conference (EUSIPCO), Kos, Greece, Aug 28 – Sep 2, 2017.