Kalman Filter for Video Object Segmentation and Tracking

Abstract

In this paper, a technique based on a discrete Kalman filter algorithm is proposed to follow the trajectory of the objects. The aim is to obtain a precise prediction of their position and motion. The accurate prediction improves both the recursive spatio-temporal segmentation and object tracking performances, enabling a high level understanding of the scene dynamics. The derived scene representation obtained finds applications in various domains. For instance, it is very well suited for dynamic scene analysis where a deep scene understanding is required. Typical examples are scene understanding and robot vision. It is also very appealing in the context-based video coding(MPEG-4). Experimental results have shown that this method is able to integrate over time the temporal information for each object and to interpolate or extrapolate its trajectory, correctly predict the position and the motion of temporal coherent objects. However, if the object has performed maneuvers, the Kalman filter fails in its prediction. In order to decide when is convenient to use the last estimated motion of the object instead of the Kalman prediction a test based on motion compensation error is used. Finally the proposed algorithm has shown its robustness in the presence of object occlusions.