Open Set Recognition for Unique Person Counting via Virtual Gates

Retail shops or restaurants are interested in real-time profiling analysis of customer visit patterns, which could enable efficient management and target marketing. They need to know not only how many people entered but also if they are visiting for the first time and keep track of their exact number. As a result, in this paper we define the new variant of unique counting for videos, that is counting new persons who have not already been counted in the past. To this end, we propose a complete real-time system which is able to perform detection, tracking and unique counting in the wild with user drawn gates. A fine-tuned network on persons body is used to extract descriptors which are more privacy-oriented. Experiments of the system on the challenging DukeMTMC dataset show that our method is able to effectively count people in real time and discern between the persons which do multiple passages through the gates.