Efficient Application of Adaptive Gaussian Kernel for Density Map Generation With Drone Data

Acquiring Crowd information using Convolutional Neural Network techniques evolved promptly. The use of drones has also increased in recent years for gathering crowd information and analysis. These algorithms or CNN models require datasets that have the ground truth available for training and testing purposes. Among the various popular crowd datasets, most datasets are captured from still street cameras or use a high building point to acquire crowd imagery. In this paper, authors introduce a new “DroneVision” crowd Dataset captured using a drone for crowd counting and mapping. Further, the lower density crowd datasets use fixed sizes Gaussian kernel to generate density maps. While the dense crowd datasets use an adaptive Gaussian kernel as the head size varies based on distance and location context of the image. Drone data is captured at various heights and distances with various crowd densities therefore suitable adaptive Gaussian kernel needs to be identified for this dataset. In this study, the authors analyze and optimize the adaptive Gaussian kernel for the generation of a density map for the introduced dataset.