Adaptive Mixtures for Object Tracking
A. Rareş, M.J.T. Reinders · 2000
In this paper we present a system that robustly tracks objects in an image sequence. The approach taken is based upon adaptive statistical models. An object selected in a frame by a user is tracked throughout the sequence by using a blob-like description of its features. The proposed coarse object description allows for a flexible representation and makes the tracking scheme robust. Tracking objects in an image sequence can principally be done either in a bottomup way or in a top-down manner. In the bottom-up way, one generally calculates the motion field between successive frames and then tries to link p ixels to each other that move similarly under the assumption of certain motion models. When approaching it in a top-down way, a model - usually a geometric description - of the object is fitted to each of the recorded frames. In this paper, we will introduce a n ob ject tracking scheme that balances between these two approaches. The objects are detected by representing them through a model. However, in stead of using a geometric description we propose to use a statistical description of the objects characteristics. Thereafter, the statistical model of the object is updated based on the observed object characteristics. This is achieved by posing the object tracking as a dynamic classification problem, i.e. for each frame the pixels are classified as to which objects they belong, an information which is then used to update the statistical models that represent these objects. In this paper, we have chosen color and position as features to discriminate between objects (the scheme, however, is generic enough to include any other features as well). The tracked objects are statistically represented as a mixture of Gaussians. In this way, coarse blob-like descriptions of the object’s shape are readily achieved. The statistical models are initialized manually by indicating the objects of interest in the first frame of a sequence to