Content-based medical video retrieval based on region motion trajectories
Résumé
In this paper, we address the problem of contentbased medical video retrieval. We propose the use of motion tracking to generate video features. First, we extract motion vectors derived from the 'MPEG-4 AVC/H.264' standard. Second, motion segmentation of the image sequence is performed by a combination of k-means clustering and motion consistency verification. Third, we used the well-known Kalman filter to track region motion between consecutive frames. This produced region's correspondences are concatenated to construct the region's motion trajectories. Finally, to compare videos, we adopted an extension of dynamic time warping (EDTW) to multidimensional time series. Results are promising: a retrieval precision at 5 of 62 % was achieved.