Il progetto
In a socially acceptable environment, understanding and analyzing users' activities and behaviors is becoming imperative. Users' behavior can provide important statistics and insights on what happens inside a given space, and which are the interactions with the objects. In retail, for instance, many efforts have been devoted toward monitoring how customers move about in this space and interact with products. Nevertheless, this challenge is still open due to different important problems, which comprise occlusions, appearance changes and dynamic and complex background. Different issues related to automating purchasing tasks as well as issues related to analytics services are currently investigated by the research community. We propose EXTRA EYE (Egocentric and eXocenTRic views for An object-level human bEhavior analYsis and undErstanding through tracking in complex spaces), a computer vision system which operates from two main points of view: an "exocentric" (third-person) point of view given by fixed cameras placed in the environment and an "egocentric" (first-person) point of view offered by wearable cameras worn by the customers. EXTRA EYE relies on three pillars, which together will contribute to Human Behavior Analysis (HBA): a First-Person View (FPV) module which aims to understand take/release user-object interactions from egocentric vision, a Third-Person View (TPV) module which processes exocentric signals, accompanied with an aggregation unit that also integrates information from the other components to perform context-specific behavioral analysis, and an Object Tracking (OT) module which supports the former two components by tracking objects of interest from both the third- and first-person points of view. For the first time, this project will consider both first-person and third-person points of view for Human Behavior Understanding in the context of human object interaction. It is worth noting that, while we choose retail stores as a clear application context, EXTRA EYE can be used in a variety of scenarios in which human behavior is largely determined by the way in which people interact with the objects available in the scene. Examples of such scenarios include factories, medical facilities, and home contexts, in which the information extracted using the proposed algorithms can be useful to track the user's behavior (e.g., to assess safety or measure quality of life), as well as to provide assistance on how to relate to specific objects (e.g., to offer information on how to use an appliance or on how to carry out a given procedure)
I risultati
Le attività di divulgazione