Grouping viewpoint images into scenes based on similarity between frames

We propose a method of automatically generating a snapshot sequence, which describes usual events, from head-mounted video camera. And we discuss a relationship between the snapshot sequence and behavior. We extract a snapshot from its video frames when the head hardly moved. This condition appears when the subject kept observation. As the eye (head) motion relates to his behavior, the snapshot sequence relates to his behavior, too. At its condition, neighbor frames are highly similar. So, we judge the head motion by similarity between neighbor frames. We use similarity estimation method by local grayvalue invariants.