This research project is concerned with the study and the development of a system for video surveillance based on multimedial biometric data. The main aspects of the project will involve:
- the identification of effective techniques for the mathematical description of multimedial data utilized in the video surveillance process;
- the identification of data mining methods based on statistical learning techniques, allowing an efficient classification of the available imagery;
- the realization of a software tool on the basis of an existing database.
In video surveillance there are operative frameworks where the entrance of individuals is allowed without any recognition, by means of an access recording; in this case people are allowed to entry after they have recorded some biometric data. An example is the admittance in bank buildings: the data are stored according to some privacy rules and can be used by authorized personnel to control ‘a posteriori’ what happened.
One of the main requirements for these control and admittance systems is to be able to retrieve information on the recorded transits starting from a specific transit. At the same time, it is important to facilitate the individuals’ access, limiting obstacles, waiting time and collaboration requirements. A system with these properties can be used also in situations whereby security barriers are not necessary.
Starting from these requirements, The goal of this research project is to study and develop a tool which is able to organize multimedial data (text, images, videos) stored during a long time, with the aim of retrieving information concerning one or more unknown individuals. To realize such a tool, we will accomplish the formulation, implementation and validation of:
- methods for image content description, ad hoc for the present problem;
- methods for data mining allowing the effective classification of available data and retrieval of elements closer, in some topological sense, to the ones given as input to the system.
In this framework, we will study local and global methods for describing the image content in a robust way with respect to variations in the environmental conditions, as, for example, wavelet descriptors, local LBP-like representation, eigenfaces representations. Furthermore, we will work at the development of semi-supervised or unsupervised learning methods. In particular we will focus on methods like spectral clustering, diffusion maps and unsupervised kernel methods.
Our technological platform will be an archival system already at disposal of this collaboration. This system has been designed in order to optimally manage a priori information on the recorded data.