Feature-based cluster segmentation of image sequences

One of the crucial points in object segmentation within image sequences is the interdependence of different features that classify some area as an object. This paper introduces a concept of cluster segmentation, which acquires different features on a pixel basis. Weighting of these features based on predefined rules is applied, in order to judge the evidence of each particular feature for the final classification. To determine the various clusters, we use a procedure which is similar to vector quantization. This allows the tracking of classification results over time, because cluster labels change only gradually from frame to frame. Furthermore, a technique for local feature analysis is applied for segment merging after global classification. The most common features used for object separation in image sequences are color and motion. The results indicate that reliable segmentation and tracking of objects can be accomplished, using this low-complexity technique