Mario Vento

Subgraph Transformations for the Inexact Matching of Attributed Relational Graphs

An inexact matching algorithm for Attributed Relational Graphs is presented: according to it, two graphs are considered similar if, by using a defined set of syntactic and semantic transformations, they can be made isomorphic to each other. The matching process is carried out by using a State Space Representation: a state represents a partial solution of the matching between the graphs, and a transition between two states corresponds to the addition of a new pair of matched nodes. A set of feasibility rules are introduced for pruning states associated to partial matching solutions which do not satisfy the required graphs morphism. Results outlining the computational cost reduction achieved by the method are given with reference to a set of randomly generated graphs.

An algorithm for recovering camouflage errors on moving people

In this paper we present a model-based algorithm working as a post-processing phase of any foreground object detector. The model is suited to recover camouflage errors producing the segmentation of an entity in small and unconnected parts. The model does not require training procedures, but only information about the estimated size of the person, obtainable when an inverse perspective mapping procedure is used. n nA quantitative evaluation of the effectiveness of the method, used after four well known moving object detection algorithms has been carried out. Performance are given on a variety of publicly available databases, selected among those presenting highly camouflaged objects in real scenes referring to both indoor and outdoor environments.

Tracking System with Re-identification Using a RGB String Kernel

People re-identification consists in identifying a person that comes back in a scene where it has been previously detected. This key problem in visual surveillance applications may concern single or multi camera systems. Features encoding each person should be rich enough to provide an efficient re-identification while being sufficiently robust to remain significant through the different phenomena which may alter the appearance of a person in a video. We propose in this paper a method that encodes peoples appearance through a string of salient points. The similarity between two such strings is encoded by a kernel. This last kernel is combined with a tracking algorithm in order to associate a set of strings to each person and to measure similarities between persons entering into the scene and persons who left it.