Face re-identification for digital signage applications

The estimation of soft biometric features related to a person standing in front an advertising screen plays a key role in digital signage applications. Information such as gender, age, and emotions of the user can help to trigger dedicated advertising campaigns to the target user as well as it can be useful to measure the type of audience attending a store. Among the technologies useful to monitor the customers in this context, there are the ones that aim to answer the following question: is a specific subject back to the advertising screen within a time slot? This information can have an high impact on the automatic selection of the advertising campaigns to be shown when a new user or a re-identified one appears in front the smart screen. This paper points out, through a set of experiments, that the re-identification of users appearing in front a screen is possible with a good accuracy. Specifically, we describe a framework employing frontal face detection technology and re-identification mechanism, based on similarity between sets of faces learned within a time slot (i.e., the models to be re-identified) and the set of face patches collected when a user appears in front a screen. Faces are pre-processed to remove geometric and photometric variability and are represented as spatial histograms of Locally Ternary Pattern for re-identification purpose. A dataset composed by different presentation sessions of customers to the proposed system has been acquired for testing purpose. Data have been collected to guarantee realistic variabilities. The experiments have been conducted with a leave-one-out validation method to estimate the performances of the system in three different working scenarios: one sample per presentation session for both testing and training (one-to-one), one sample per presentation session for testing and many for training (oneto- many), as well as considering many samples per presentation sessions for both testing and training (many-to-many). Experimental results on the considered dataset show that an accuracy of 88.73% with 5% of false positive can be achieved by using a many-to-many re-identification approach which considers few faces samples in both training and testing.