Single camera based visual tracking and multilayered scene modelling

Abstract

The most interesting information in video images is often related to moving objects. In tracking applications, one follows moving objects or people throughout a sequence of images. The system determines the position of the object of interest in each frame, which results in a track. Applications are often found in traffic, surveillance, sports analysis.Occlusion is an import issue in all tracking systems. In this case, interesting objects are partially or fully hidden behind other static or moving objects. Camera’s which can identify these occluded objects, will generate less faulty information. Consequently, tracking and occlusion are strongly related to 3D scene models.My research is focused on the development of a robust, self learning, single camera tracking system, which is based on an automatically built, multilayered scene model. Single camera systems cannot directly provide depth accurate information, but they can still gather a lot of useful depth knowledge about a scene, based on perspective effects, known planes (for example, the floor) and the analysis of occlusions. This requires a scene model which tells the user where the occluding objects are, where people can walk and cars can drive…

Publication
14th UGent FEA PhD Symposium