Action Recognition in Meeting Scenarios using Global Motion Features

Abstract

Feature Extraction An approach for an off-line segmentation- and recognition system for actions in meeting scenarios is presented. The deployed system consists of three major modules, which are a feature extraction module, a segmentation module and a statistical classifier which all have been assembled in conjunction of an EU funded automated meeting transcription system. All these modules and their functionality are briefly introduced. The performance of the system is evaluated on a part of the PETS-ICVS 2003 dataset (scenario B, camera 1 and 2). It is shown, that the presented system is able to detect and recognize single person actions in a meeting with an acceptable performance. One feature stream per person