Human crowd detection for drone flight safety using convolutional neural networks

In this paper a novel human crowd detection method, that utilizes deep Convolutional Neural Networks (CNN), for drone flight safety purposes is proposed. The aim of our work is to provide light architectures, as imposed by the computational restrictions of the application, that can effectively distinguish between crowded and non-crowded scenes, captured from drones, and provide crowd heatmaps that can be used to semantically enhance the flight maps by defining no-fly zones. To this end, we first propose to adapt a pre-trained CNN on our task, by totally discarding the fully-connected layers and attaching an additional convolutional one, transforming it to a fast fully-convolutional network that is able to produce crowd heatmaps. Second, we propose a two-loss-training model, which aims to enhance the separability of the crowd and non-crowd classes. The experimental validation is performed on a new drone dataset that has been created for the specific task, and indicates the effectiveness of the proposed detector.