A Robust and Efficient Overhead People Counting System for Retail Applications

Counting people with overhead cameras is a feature that is increasingly required by the retail market, for surveillance and business intelligence. However, despite the great advances in modern neural networks, it is far from simple to train effective systems in all possible real-world scenarios. The main problem is that the publicly available datasets are not sufficiently representative or are not annotated for this purpose. To this aim, in this paper we demonstrate that a considerable effort in the collection of heterogeneous data in real scenarios, producing a new dataset of about 30,000 images, allowed to realize an effective and efficient people counting system, able to process at least 10 FPS on board of three different types of smart cameras. In addition, the F-Score higher than 95% over the test set demonstrates the effectiveness of the proposed people counting system and its suitability for real applications.