In this project we will explore the use of sparse techniques to improve the estimation of multiple motion fields as well as spacevarying switching matrix (stochastic matrix) that describes the switching process associated to the movement of each target. As scientific outcomes of the project, we expect to obtain reliable estimates for the model parameters (many hundreds), to remove the over smoothed character of the field estimates. We also expect to speed up the estimation procedure bringing it closer to real time applications and extend these techniques to multicamera settings. We also expect to develop an adaptive algorithm for the online estimation of multiple motion fields able to update in a recursive way the number of fields and the field estimates whenever new information arrives.

The algorithms developed in this project will be applied to the classification of human activities using outdoor video data obtained in previous projects (CAVIAR, URUS, A second direction will be the application of these techniques to multicamera surveillance systems which raise several new problems related to image registration, cooccurence and data fusion. We will also apply the model and the methods developed to the related problem of nonlinear system identification. We believe the proposed model will be an important tool for the analysis of motion in video signals and it is not restricted to human activity recognition.

The project team has strong experience in both pillars of this research proposal. In fact most of its members participated in the FCTARGUS project in which the multiple motion field model was developed (see http://users.isr.ist.utl.pt/~jsm/ ARGUS/). The team includes also specialists in sparse systems and in nonlinear dynamical systems and all the team members (6 with PhD) have long experience in the coordination and participation in projects, both at national and European levels.