MUBAI: multiagent biometrics for ambient intelligence

Present technological progress opens new scenarios, where people can interact with electronic equipment which is embedded in every-day objects and settings. The so called ambient intelligence allows to automatically detect context from wearable or environmental sensor systems and to exploit such information for adaptive support. Domotic systems are a spreading example of such philosophy. Advanced sensors and devices make up context-aware environments which are sensitive and responsive to the presence of users. When each single user in a set can be recognized, it is possible to provide enhanced personalization of functionalities and services. Recognition in such situations is preferably not bound to a voluntary, or conscious at least, user’s interaction with the equipment, but rather relies on the ability of an underlying control system to automatically and autonomously catch some user’s characteristic and to use it for identification. This implicit requirement makes reference to biometric techniques. We propose here Multiagent Biometrics for Ambient Intelligence (MUBAI) architecture, which specifies the composition of more biometric modules in a multiagent recognition system. In principle, each module implements an autonomous agent, which performs its own recognition process; however, not all such agents are equally reliable on any single input. System Response Reliability (SRR) allows to assess reliability on a single-response basis, and is a crucial element during fusion of agents’ results. A further improvement is obtained by communication/collaboration activities, which are a core characteristic of multiagent systems. MUBAI architecture exploits two types of agents, according to the “the brawn and the brains” approach: more Classifier Agents perform biometric processing on (possibly) different traits exploiting the inter-agent communication ruled by the N-Cross Testing Protocol. A different type of agent implements a Supervisor Module to produce a final recognition result and to possibly update Classifiers’ parameters. Such agent allows overcoming the parameter invariance to which present multibiometric architectures are bound. In the experiments presented in this paper, the assessed MUBAI instance includes four modules implementing different face recognition techniques, and a supervisor module. The obtained results demonstrated how MUBAI allows to achieve far better performances than single classifiers. As a further contribution, we show how MUBAI architecture can also be used in a dynamic setting, where new agents added from time to time to the system can be effectively trained online.