FUSION FRAMEWORK FOR VIDEO EVENT RECOGNITION

摘要

This paper presents a multisensor fusion framework forrnvideo activities recognition based on statistical reasoningrnand D-S evidence theory. Precisely, the frameworkrnconsists in the combination of the events’ uncertaintyrncomputation with the trained database and the fusionrnmethod based on the conflict management of evidences.rnOur framework aims to build Multisensor fusionrnarchitecture for event recognition by combining sensors,rndealing with conflicting recognition, and improving theirrnperformance. According to a complex event’s hierarchy,rnPrimitive state is chosen as our target event in thernframework. A RGB camera and a RGB-D camera arernused to recognise a person’s basic activities in the scene.rnThe main convenience of the proposed framework is thatrnit firstly allows adding easily more possible events intornthe system with a complete structure for handlingrnuncertainty. And secondly, the inference of Dempster-rnShafer theory resembles human perception and fits forrnuncertainty and conflict management with incompleterninformation. The cross-validation of real-world data (10rnpersons) is carried out using the proposed framework, andrnthe evaluation shows promising results that the fusionrnapproach has an average sensitivity of 93.31% and anrnaverage precision of 86.7%. These results are better thanrnthe ones when only one camera is used, encouragingrnfurther research focusing on the combination of morernsensors with more events, as well as the optimization ofrnthe parameters in the framework for improvements.