An Extended Snake Model For Real-Time Multiple Object Tracking

In this paper we present a snake model dedicated to real-time multiple object tracking. This model is able to track moving non-rigid objects in video sequences obtained from moving camera. Contrary to other approaches, here the snake deformation is performed without any preprocessing and the tracking method does not rely on any motion estimation step. The tracking method is based on two simple successive steps which are snake initialisation and snake deformation. Avoiding complex additional steps allows us to propose a tracking method which can be performed in real-time. Two specific extensions of classical applications are proposed in this paper. The first one gives the snake the ability to divide in several contours if necessary. It is particulary useful when two objects which were close together move apart. It also improves the robustness of the tracking method against noise characterized by high gradient values. The second extension deals with multiresolution processing of the image. In order to limit computation time, we propose to process low-resolution frames instead of original ones. A quality criteria is then used to determine if the result obtained is correct or not. In the latter case, the current frame is analysed at a higher resolution. The proposed extended snake model has been successfully applied to real-time multiple object tracking of football players in football game video sequences. Players positions obtained are then further used in a real-time football video analysis tool.