We propose a shot segmentation method of live video stream for wearable assistants. This method can automati-cally divide an unedited video stream into meaningful shots, even if the stream contains shaking, vibration or blurring. The basic idea of the method is dividing video streams into "stationary" shots and "transitive" shots based on the tendency of visual changes. Two typical conventional methods were compared with our method on the 55 minute maintenance video recorded in a restricted environment. Analyzing the video and computing the recall and precision rate showed the adequacy of our method. Furthermore, the practical effectiveness of the methods was evaluated by another 54 minute patrol video in an outdoor environment. Nine human subjects tried to find randomly selected shots with different segmentation produced by the three methods. The result shows that our method supported users to find the shots most effectively (89% success in 5 minutes).
展开▼