High and low level object descriptions for video tracking process

In this paper a new segmentation algorithm approach for real time traffic scenes is proposed, combining high level and low level object descriptions. Both descriptions make it possible to develop a tracking method, robust regarding occlusions, region clustering and brightness variations. High level description is defined by geometric attributes and motion model. Updating these features (associated to each object) can be obtained by a low level segmentation which is based on a background update approach, associated with a spatial-temporal segmentation. This spatial-temporal segmentation is built on a motion estimation taken out from a modified Expectation-Maximization (EM) method. These two descriptions leads to a really efficient strategy in terms of robustness, over or sub-segmentations and occlusions. Furthermore, under severe brightness changes, our new temporal algorithm also permits a perfect background update control. Some real traffic examples are included at the end of this paper.