Object tracking using a combined appearance and geometric model
Pierre F. Gabriel, Justus Piater, Jacques Verly · Open Repository and Bibliography (University of Liège) · 2004
There has been considerable research activity in the area of objects tracking in video sequences over the last 20 years. However, most current tracking systems do not deal well with complex situations, where objects can have different sizes, be of similar appearances or occlude each other. In fact, most of them are based on either a region extraction with a background subtraction or a contour extraction using particle filtering. The main goal of this work is to track objects from image to image in a video sequence. More precisely, we want to track each interest point associated to an object. In fact, to do this, we apply the color Harris detector [2] to the region of interest (ROI) of each object and we search for a correspondence between the interest points in the current frame and those in the previous frame based on both an appearance model and a geometric model. The appearance model consists of characterizing each detected interest point with a vector containing some local attributes. After a study of different configurations of characterizations, we chosen the color local jet [3] up to order 1 and added the cornerness (i.e. the value given by the color Harris detector) to the vector. The geometric model consists of two geometric parameters: the distance from the