Hierarchical Correspondance between Physical Situations and Action Models
|
Abstract:
The framework of the paper is the recognition of ongoing actions of groups of agents in a scene watched by image sensors. In such applications, the actions are not directly observed but must be interpreted from observed physical situations. In particular, as the perception ressources are usually limited, one crucial issue is whether more informed situations correspond to more specific actions. The situations are given an extensional representation and partially ordered w.r.t some information specificity. The actions are represented using Description Logics, thus organized into subsumption taxonomies. The correspondance between both representation spaces is formalized and shown to behave monotonically w.r.t. the respective taxonomic structures under specific "homogeneity " conditions over the situations. The application of this formal framework to situation recognition is then discussed.
