My current activities focus on computational aspects related to the analysis of the visual scene. In particular, I am interested in the study of principled Bayesian approaches built upon models that attempt to mimic the physical processes behind measurement formation and scene dynamics. In the context of a multibody tracking problem, such models may e.g. implement the occlusion process, or the motion model of an articulated body that obeys physiological constraints. While powerful in principle, such approaches are usually demanding in terms of computations, and remain often not applicable in practice. For the multibody tracking problem, whose principled formulation induces an estimation algorithm with exponential complexity, [1] proposes an approximate framework which can handle occlusions and avoids track coalescence at an affordable (sub-quadratic) computational cost. In conjunction with [2], it defines a new class of efficient tracking filters able to self-manage the trade-off between computational efficiency and robustness (see SmarTrack section below). Future work will aim at integrating complementary sensing (e.g. acoustic, infrared, ...), extending it to handle articulated bodies for supporting motion capture and gesture recognition, and to apply it in different contexts (e.g. vehicle tracking, ...). Also, tracking in complex environments (e.g. a museum) with the need of distributed sensing and processing will be addressed. Envisaged application contexts include smart environments, sports analysis, and surveillance.

SmarTrack - a SmarT people Tracker. Based on the work published in my PAMI paper of 2006, the prototype people tracker SmarTrack has been realized. SmarTrack has been showcased at the CHIL exhibition stand during the IST 2006 event, which earned one of three "Best IST'06 Exhibit" award. SmarTrack monitored continuously for three days the CHIL EU project booth, computing more than 900 accurate 3D tracks. (Smar)Tracks have been used to provide a service to the booth visitors: an automatically generated personalized report has been handed out to each visitor at the booth exit containing a plot of his path through the booth and a brief description of the demo he was most interested in (i.e. where, according to SmarTrack, he has spend most of the time). The prototype now tracks up to ten people in real time through multiple persistent occlusions in cluttered environment. [http://tev.fbk.eu/smartrack/]