Detecting scenes of attention from personal view records - motion estimation improvements and cooperative use of a surveillance camera
Satoshi KUBOTA, Yuichi Nakamura, Yuichi Ohta · 2002
This paper introduces a novel method for analyzing video records captured by a head-mounted camera. Compared to our previous method, the new method is improved on two points. One is a new method of two-step motion estimation that adaptively uses either of the 2D affine model or the 3D rigid-body motion with central projection model. The other is a cooperative use of a wide-angled surveillance camera, which delineates the location and the situation where the Videos captured by a head-mounted camera (here after abbreviated as HMC) are good media for recording our activities, and they are useful for recalling or sharing the experience afterwards. Videos taken as personal records. however. can be long and redundant, and a user may n&d considerable tim;for finding the i'nformation he/she requires. This disadvantage may spoil the merit of video records. For this purpose, we previously reported that scenes of attention can be good indices for summarizing those videos[3]. The view from an HMC contains the central portion of the sight, and the camera's ego-motion represents the user's head motion. By estimating egomotions and by separating object motions, we can detect typical behaviors of the user's for paying attention to something as shown in Fig. 1. Figure 2 shows a browser that presents those scenes, and this browser is much more comprehensible than a simple arrangement of images taken at regular intervals (Fig. 3). We also reported that video summaries composed of those scenes showed good match to the summaries that were manually made by selecting important scenes from the videos[4]. This paper introduces two new approaches for the performance improvement and for the extension of the potential applications. mwitlg