Digital Scene Augmentation Techniques for Generating Photo-Realistic Virtual Crowds

Crowd estimation has a wide range of applications especially in relation to computer vision, robotics and security surveillance technology. New computer vision techniques and deep learning technology have enabled large scale crowd estimation but advances in these techniques have been hindered by the lack of high quality, annotated and publicly available datasets. Although there has been several attempts to compile crowd datasets in the past, collecting and labelling the data is a tedious and labour intensive task. New privacy legislations also make it difficult to release real world footage to the public. In this paper, we present a novel method to generate photo-realistic scalable labelled synthetic crowds for the purpose of accelerating the state-of-the-art in crowd understanding techniques. We generate human models on scene reconstructed environments. The environments are created with footage captured by aerial drone surveys. The crowds are then compo sited with the original images to generate photo-realistic data. This dataset contains 500 high resolution images with over 230,000 annotations and is intended to be publicly available to further advance research in crowd understanding.