Tracking abnormalities in video capsule endoscopy using surrounding features with a triangular constraint

This paper proposes a method to track abnormalities in successive frames in a capsule endoscopic image sequence. Exact tracking of an abnormality in the gastrointestinal tract is useful in preparing the content for educational systems. However, if the abnormality is de-formable over continuous frames and its features are not highly distinct, it is difficult to track abnormalities precisely. The proposed method uses not only the abnormality image features, but also surrounding features, called supporters. In a capsule endoscopic image sequence, however, both the supporters and the target are difficult to track using only their image features, because these features are indistinguishable. Since estimation using only surrounding features is hard, the proposed method uses triangular constraints among supporters as well. The proposed method is able to track successfully even if the target and supporter features are indistinguishable. Finally, we evaluate the proposed method using eight major types of abnormalities.