Real Time Multiple Object Tracking Based on Active Contours

Abstract

In this paper our purpose is to present some solutions to multiple object tracking in an image sequence with a real time constraint and a possible mobile camera. We propose to use active contours (or snakes) modelling. Classical active contours fail to track several objects at once, so occlusion problems are difficult to solve. The model proposed here enables some topology change for the objects concerned. Indeed a merging and a splitting phases are respectively performed when two objects become close together or move apart. Moreover, these topology changes help the tracking method to increase its robustness to noise characterized by high gradient values. In the process we have elaborated, no preprocessing nor motion estimation (which are both time consuming tasks) is required. The tracking is performed in two steps that are active contour initialisation and deformation. The process supports non-rigid objects in colour video sequences from a mobile camera. In order to take advantage of compressed formats and to speed up the process when possible, a multiresolution framework is proposed, working in the lowest-resolution frame, with respect to a quality criterion to ensure a satisfying quality of the results. The proposed method has been validated in the context of real time tracking of players in soccer game TV broadcasts. Player positions obtained can then be used in a real time analysis tool of soccer video sequences.