Video Analytics for Indoor Crowd Estimation

Abstract

This paper demonstrates the usage of security camera footages with deep convolutional neural networks to provide cabin-level crowd density estimates in the video frames. Some applications for this include cabin-level crowd density estimates of incoming trains. With this information, train passengers may choose to board the trains at less crowded cabins, potentially decreasing the dwell time of trains at stations and experiencing a more pleasant commute overall. In a way, the crowd level estimation information will also help to maximize the train and platform capacity. Leveraging on the security camera footages would also serve as a cost-effective solution to the train operator as compared to installing new sensing equipment in the trains. Due to privacy and security concerns of publishing train cabin video frames, this paper will present the experiment results on an indoor pedestrian dataset.